Tiến độ hàng đợi

Đếm mỗi run_id một lần, không đếm từng file job: run_id đã ok ở bất kỳ đâu (kể cả job bị từ chối result_conflict vì kết quả ok đã có) tính là xong một lần, các bản xếp lại dưới manifest khác không cộng thêm. Job chưa có trạng thái mà ghim code MỚI hơn bản đang publish cho worker là "chờ publish": sẽ chạy sau lần publish kế tiếp nên vẫn tính vào queued, tổng và ETA. Job ghim code CŨ hơn (hoặc rẽ nhánh khác) là "bị thay thế (code cũ)": không worker nào chạy nên không tính vào tổng. Trạng thái running quá thời gian thường lệ của job (trung vị các lần chạy xong, mặc định 6h 00m) mà không còn worker giữ claim là "treo": lượt worker sau sẽ chạy lại. ETA = (queued + running + treo) × trung vị thời gian các run đã xong trên máy đó (không có ETA khi chưa run nào xong); ok gồm cả run đã ingest. Mỗi máy tách thanh theo phiên bản nhãn. Cập nhật lúc 2026-10-25 15:35 UTC.

zephyrus · 242/243 đã chạy xong · 11 bị thay thế (code cũ)

m1: 187/187

v1: 49/50

v2: 1/1

v2_1: 5/5

queued 0 · running 0 · treo 1 · ok 235 · failed 7 · ETA ~7m

treo (worker không còn chạy): e2_mmbert_base_cross_ce_real_frozenemb_lr1e5 seed 42 trên zephyrus, bắt đầu 2026-10-22 04:55 UTC (đã 82h 40m)

legion · 70/70 đã chạy xong · 56 bị thay thế (code cũ)

m1: 66/66

v1: 4/4

queued 0 · running 0 · treo 0 · ok 69 · failed 1

kaggle · 7/7 đã chạy xong · 6 bị thay thế (code cũ)

m1: 7/7

queued 0 · running 0 · treo 0 · ok 7 · failed 0

Giải thích nhãn

6 intent (thứ tự ưu tiên)

Khi nhiều intent cùng khớp, intent đứng trước thắng: omi_intent_repeat > callback > relationship > nlu_fallback > negate > affirm. Ví dụ bên dưới do người viết tự đặt, không lấy từ data thật.

Nhãn	Tên	Định nghĩa	Ví dụ (bot → khách)
`omi_intent_repeat`	Yêu cầu nhắc lại	Khách muốn bot nhắc lại (“nói lại”, “gì cơ”, “nghe không rõ”), hoặc “alo” khi đang chờ người nghe hoặc không nghe rõ.	Bot: “Hợp đồng của anh sẽ đến hạn vào ngày mười lăm.” Khách: “Chị nói lại giúp tôi với.”
`callback`	Hẹn gọi lại	Khách hay người nghe máy bảo gọi lại, hẹn gọi lại hoặc cho số khác để liên hệ. Chỉ báo bận hoặc chỉ nêu giờ mà không bảo gọi lại thì không phải callback.	Bot: “Em xin phép trao đổi vài phút được không ạ?” Khách: “Giờ anh đang bận, chiều gọi lại nhé.”
`relationship`	Người thân, không phải chủ hợp đồng	Người nghe máy không phải người bot cần gặp và nói rõ mình là ai (người nhà, bạn, đồng nghiệp, “nghe máy giùm”). Chỉ báo người đó vắng mà không xưng mình là ai thì không phải relationship.	Bot: “Em muốn gặp anh Nam, anh có phải anh Nam không ạ?” Khách: “Anh là em trai, anh ấy đi vắng rồi.”
`nlu_fallback`	Không hiểu / lạc đề	Câu khách không trả lời điều bot hỏi: lạc đề, câu méo do ASR, hỏi ngược, phàn nàn, nghi lừa đảo, hỏi về tiền hay thủ tục; kể cả khi mở đầu bằng “dạ / ừ” rồi nói chuyện khác, hoặc chỉ có từ đệm khi bot đang xin một thông tin cụ thể.	Bot: “Anh xác nhận giúp em thông tin hợp đồng nhé?” Khách: “Mà tháng này phải đóng bao nhiêu tiền vậy em?”
`negate`	Phủ nhận	Khách phủ nhận, không đồng ý hoặc sửa điều bot nói (không phải người đó, nhầm số, chưa nhận, sai thông tin); không nhớ / không muốn cho thông tin; hỏi xác nhận để làm gì; bận hay vắng mà không bảo gọi lại.	Bot: “Thông tin hợp đồng của mình có chính xác không ạ?” Khách: “Không, sai rồi em ơi.”
`affirm`	Đồng ý	Không rơi vào 5 nhãn trên và câu có từ xác nhận rõ điều bot hỏi (“đúng rồi”, “có”, “nhận rồi”, “tôi nghe đây”); “dạ / ừ” đứng một mình chỉ là affirm khi bot hỏi có/không hoặc hỏi danh tính.	Bot: “Anh là chủ hợp đồng đúng không ạ?” Khách: “Dạ đúng rồi em.”

Ranh giới hay nhầm — quy tắc v2 (data lv2)

Khách hỏi “để làm gì” khi đang được xác minh thông tin → negate.
Khách hỏi về tiền hoặc thủ tục mà không có từ đồng ý hay phủ nhận → nlu_fallback.
Khách bảo gọi lại, hẹn gọi lại hoặc cho số khác → callback; chỉ báo bận hoặc chỉ nêu giờ → negate (theo spec).
Chỉ “ừ”, “dạ” sau lời bot thông báo (không hỏi) → affirm. Khác spec (spec gán fallback) vì train và QC đều gán affirm.

m1: bộ dữ liệu chính, do khách cập nhật thủ công: tập dòng, nhãn, dev và test đều riêng nên không so trực tiếp với v1. Đầu vào test là chữ ASR lấy từ file QC gốc. Các dòng QC khách đã sửa nhãn so với bản gốc được tách khỏi bảng chính vì nghiêng về model (phần lớn trùng dự đoán của model), chỉ hiện ở split contrast để chẩn đoán. Không test_informed.
v2: Nhãn QC v2 được chọn sửa một phần từ các dòng mà model bỏ phiếu khác nhãn gốc → số test v2 thiên về model, chỉ để tham khảo, không dùng để chọn phương pháp.
v2_1: nhãn train/dev sửa theo quy tắc 1 (theo guideline v2, chủ dự án chọn) và quy tắc 3 (theo spec), dev giữ đúng các dòng của v1; nhãn test giữ nguyên v1 → so trực tiếp được với baseline.

Các split

dev: tập tách từ train, dùng chọn ngưỡng và xếp STT.
test_clean: QC thật, đã bỏ cặp bot + khách trùng train; bảng chính.
test_reply_unseen: QC thật, câu khách chưa từng xuất hiện trong train; mức chặt nhất.
test_full: toàn bộ QC, gồm cả cặp trùng train nên số liệu lạc quan; chỉ để đối chiếu.
contrast: chỉ có ở m1: các dòng QC khách đã sửa nhãn so với bản gốc. Nghiêng về model nên chỉ để chẩn đoán, hiện cuối trang run; không vào bảng, không xếp hạng, không chọn ensemble.

Cách đọc cột

Số sau ±: độ lệch chuẩn mẫu (ddof=1) giữa các seed của cùng phương pháp, tính bằng điểm phần trăm. Ví dụ 84.10 ±0.37 = trung bình 84.10 qua các seed, các seed thường lệch nhau khoảng 0.4 điểm. Đây là độ dao động do seed, không phải khoảng tin cậy trên test: với cỡ test hiện tại, chênh vài điểm vẫn có thể là nhiễu (xem bootstrap). Ô không có ± là run 1 seed hoặc baseline.
Pre / Rec / F1: precision, recall, F1 của từng intent trên split đó.
Macro: trung bình đều các intent; Weighted là trung bình theo số mẫu mỗi intent.
▲ (chữ xanh): ứng viên vượt baseline cùng bản dữ liệu nhưng chưa tốt nhất cột.
Ô nền xanh đậm: giá trị tốt nhất cột, tính trong cùng bản dữ liệu (m1, v1, v2 hoặc v2_1). Dòng transductive hoặc test_informed vẫn hiện nhưng không nhận ô xanh hay ▲.
STT 0: baseline prod, ghim đầu bảng; ứng viên xếp từ 1, riêng cho từng bản dữ liệu.
Quy tắc giải mã: cột không tiền tố dùng quy tắc mặc định: ngưỡng fallback chọn trên dev (recall cao nhất với precision ≥ 0.85). Cột argmax là cùng xác suất đó giải mã bằng argmax, không ngưỡng fallback. Baseline prod chỉ có nhãn cứng nên hai quy tắc cho cùng một số.
−rep: macro trên 5 intent, bỏ omi_intent_repeat: test_clean chỉ có vài dòng repeat nên một dòng lật làm macro-F1 đổi vài điểm.

Xếp hạng phương pháp

Hai bảng chính chỉ đo trên phần test QC model chưa từng thấy: clean = test_clean (bỏ cặp bot+KH trùng train), reply-unseen = test_reply_unseen (câu KH không có trong train, mức chặt nhất). Lựa chọn dựa trên dev + test. STT xếp theo dev recall@p85; ngưỡng và temperature chỉ khớp trên dev, không bao giờ chỉnh theo test. Số liệu là trung bình ± độ lệch chuẩn qua seed. Khi tỉ lệ lớp ở test khác dev, precision test có thể thấp hơn 0.85. Hàng tham chiếu (STT 0; nhãn cứng không có xác suất nên recall@p85 là recall fallback thường) luôn ghim ở đầu cả hai bảng, kể cả khi sắp xếp cột; ứng viên đánh STT từ 1. Bảng cuối trang là dev và test_full, dữ liệu liên quan train, chỉ để đối chiếu. Cột Data: v1 là nhãn gốc, v2 là nhãn train và QC đã relabel (method đuôi _lv2), v2_1 là nhãn train/dev theo spec với QC giữ nguyên v1 (đuôi _lv21); STT, baseline và ô tô đậm tính riêng trong từng bản, dùng bộ lọc Data để xem một bản. Tạo lúc 2026-10-25 15:35 UTC.

STT	Model_Name	Backbone	Threshold	Data	Họ	Máy	Seed	n_eff	Params	Latency ms	clean Accuracy	clean Macro_Pre	clean Macro_Rec	clean Macro_F1	clean Macro_F1 −rep	clean Weighted_F1	clean FB recall	clean FB precision	clean argmax Macro_F1	clean argmax Macro_F1 −rep	clean argmax FB recall	clean argmax FB precision	reply-unseen Accuracy	reply-unseen Macro_Pre	reply-unseen Macro_Rec	reply-unseen Macro_F1	reply-unseen Macro_F1 −rep	reply-unseen Weighted_F1	reply-unseen FB recall	reply-unseen FB precision	reply-unseen argmax Macro_F1	reply-unseen argmax Macro_F1 −rep	reply-unseen argmax FB recall	reply-unseen argmax FB precision	clean fb→negate	clean fb→affirm	clean false fb	Trạng thái	Phân tích
0	e0_prod_sft_hard_none_qc_v16_m1 tham chiếu · provenance: unknown	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	m1	reference	zephyrus	1	1	2.0 B (reported)	~1,715 ms (reported)	95.00%	89.22%	92.22%	90.16%	91.65%	94.96%	73.08%	87.16%	90.16%	91.65%	73.08%	87.16%	94.59%	86.77%	92.90%	88.63%	92.24%	94.59%	77.05%	88.68%	88.63%	92.24%	77.05%	88.68%	16.92%	7.69%	0.63%	ok	–
0	e0_prod_sft_hard_none_qc_v16 tham chiếu · provenance: unknown	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v1	reference	zephyrus	1	1	2.0 B (reported)	~1,715 ms (reported)	93.71%	86.69%	90.36%	87.68%	89.84%	93.65%	70.41%	83.13%	87.68%	89.84%	70.41%	83.13%	93.55%	86.21%	91.91%	88.04%	91.10%	93.55%	78.16%	85.00%	88.04%	91.10%	78.16%	85.00%	17.35%	9.18%	0.92%	ok	–
0	e0_prod_sft_hard_none_qc_v16_lv2 tham chiếu · provenance: unknown	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2	reference	zephyrus	1	1	2.0 B (reported)	~1,715 ms (reported)	93.58%	83.83%	90.99%	85.94%	89.80%	93.57%	70.33%	77.11%	85.94%	89.80%	70.33%	77.11%	93.32%	83.21%	92.18%	85.74%	90.88%	93.37%	76.83%	78.75%	85.74%	90.88%	76.83%	78.75%	17.58%	8.79%	1.24%	oktest_informed	–
0	e0_prod_sft_hard_none_qc_v16_lv21 tham chiếu · provenance: unknown	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2_1	reference	zephyrus	1	1	2.0 B (reported)	~1,715 ms (reported)	93.71%	86.69%	90.36%	87.68%	89.84%	93.65%	70.41%	83.13%	87.68%	89.84%	70.41%	83.13%	93.55%	86.21%	91.91%	88.04%	91.10%	93.55%	78.16%	85.00%	88.04%	91.10%	78.16%	85.00%	17.35%	9.18%	0.92%	ok	–
1	e3_ensemble_devtop1_familygeomean_m1	–	0.051	m1	ensemble	zephyrus	1	1	N/A	271.8 ms	93.00%	86.10%	89.90%	86.82%	88.53%	93.57%	97.69%	53.14%	89.62%	92.37%	85.38%	83.46%	91.86%	82.25%	85.61%	82.62%	88.23%	92.53%	98.36%	52.40%	86.77%	92.36%	86.07%	83.33%	1.54%	0.77%	5.03%	okp85_not_met	–
2	e3_ensemble_devtop1_familygeomean_tempdev_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	ensemble	zephyrus	1	1	N/A	271.8 ms	93.30%	86.38%	90.18%	87.19%	88.97%	93.80%	97.69%	54.74%	89.17%	92.33%	85.38%	84.09%	92.17%	82.49%	85.89%	82.97%	88.65%	92.77%	98.36%	53.81%	86.19%	92.32%	86.07%	84.00%	1.54%	0.77%	4.71%	okp85_not_met	–
3	e3_ensemble_devtop1_familygeomean_temphard_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	ensemble	zephyrus	1	1	N/A	271.8 ms	93.30%	86.38%	90.18%	87.19%	88.97%	93.80%	97.69%	54.74%	89.17%	92.33%	85.38%	84.09%	92.17%	82.49%	85.89%	82.97%	88.65%	92.77%	98.36%	53.81%	86.19%	92.32%	86.07%	84.00%	1.54%	0.77%	4.71%	okp85_not_met	–
4	e3_ensemble_devtop1_familymean_m1	–	0.112	m1	ensemble	zephyrus	1	1	N/A	271.8 ms	93.43%	86.98%	91.65%	88.33%	89.33%	93.89%	96.92%	55.75%	89.41%	92.12%	83.08%	83.08%	92.33%	83.95%	88.75%	85.29%	89.02%	92.88%	97.54%	54.84%	86.55%	92.09%	83.61%	82.93%	0.77%	2.31%	4.49%	okp85_not_met	–
5	e3_ensemble_sweep_k2_m1	–	0.110	m1	ensemble	zephyrus	1	1	N/A	106.3 ms	94.66%	87.61%	91.52%	89.14%	90.97%	94.81%	93.85%	68.16%	89.81%	92.38%	83.85%	84.50%	93.82%	84.04%	88.71%	85.88%	90.74%	94.02%	94.26%	67.25%	86.49%	92.36%	84.43%	84.43%	1.54%	3.85%	2.56%	okp85_not_met	–
6	e3_ensemble_sweep_k3_m1	–	0.051	m1	ensemble	zephyrus	1	1	N/A	45.3 ms	93.77%	84.79%	89.90%	86.54%	89.45%	94.11%	95.38%	59.33%	87.46%	91.15%	83.85%	77.86%	92.74%	80.11%	85.64%	81.97%	89.13%	93.17%	95.90%	58.21%	83.74%	91.08%	84.43%	77.44%	1.54%	1.54%	3.82%	okp85_not_met	–
7	e3_ensemble_sweep_k4_m1	–	0.058	m1	ensemble	zephyrus	1	1	N/A	94.7 ms	93.89%	85.07%	90.09%	86.77%	89.73%	94.25%	95.38%	59.05%	87.72%	91.47%	83.08%	78.83%	92.84%	80.35%	85.82%	82.17%	89.37%	93.30%	95.90%	57.64%	84.01%	91.40%	83.61%	78.46%	1.54%	1.54%	3.86%	okp85_not_met	–
8	e3_ensemble_sweep_k5_m1	–	0.111	m1	ensemble	zephyrus	1	1	N/A	277.3 ms	94.95%	88.65%	92.27%	89.97%	91.29%	95.13%	96.15%	67.20%	90.27%	92.32%	84.62%	83.33%	94.23%	85.70%	89.50%	87.08%	91.16%	94.44%	96.72%	66.67%	87.18%	92.31%	85.25%	83.20%	1.54%	1.54%	2.74%	okp85_not_met	–
9	e3_ensemble_sweep_k6_m1	–	0.087	m1	ensemble	zephyrus	1	1	N/A	277.3 ms	94.61%	89.32%	92.08%	90.04%	90.66%	94.86%	96.15%	63.45%	90.27%	92.32%	84.62%	83.33%	93.77%	87.30%	89.27%	87.51%	90.46%	94.06%	96.72%	62.77%	87.18%	92.31%	85.25%	83.20%	1.54%	1.54%	3.23%	okp85_not_met	–
10	e3_ensemble_sweep_k7_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.087	m1	ensemble	zephyrus	1	1	N/A	277.3 ms	94.61%	89.32%	92.08%	90.04%	90.66%	94.86%	96.15%	63.45%	90.27%	92.32%	84.62%	83.33%	93.77%	87.30%	89.27%	87.51%	90.46%	94.06%	96.72%	62.77%	87.18%	92.31%	85.25%	83.20%	1.54%	1.54%	3.23%	okp85_not_met	–
11	e3_ensemble_sweep_k8_m1	–	0.087	m1	ensemble	zephyrus	1	1	N/A	277.3 ms	94.61%	89.32%	92.08%	90.04%	90.66%	94.86%	96.15%	63.45%	90.27%	92.32%	84.62%	83.33%	93.77%	87.30%	89.27%	87.51%	90.46%	94.06%	96.72%	62.77%	87.18%	92.31%	85.25%	83.20%	1.54%	1.54%	3.23%	okp85_not_met	–
12	e3_ensemble_top_geomean_m1	–	0.073	m1	ensemble	zephyrus	1	1	N/A	149.4 ms	93.68%	85.21%	91.80%	87.63%	89.77%	94.07%	95.38%	58.22%	88.48%	92.42%	85.38%	82.22%	92.58%	81.30%	88.90%	84.02%	89.39%	93.09%	95.90%	56.80%	85.34%	92.40%	86.07%	82.03%	1.54%	2.31%	3.99%	okp85_not_met	–
13	e3_ensemble_top_mean_m1	–	0.073	m1	ensemble	zephyrus	1	1	N/A	149.4 ms	92.49%	84.43%	91.33%	86.51%	88.43%	93.20%	96.15%	50.00%	88.48%	92.43%	83.08%	83.08%	91.30%	80.66%	88.41%	82.98%	88.14%	92.14%	96.72%	49.37%	85.35%	92.41%	83.61%	82.93%	1.54%	1.54%	5.61%	okp85_not_met	–
14	e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	decoder	legion	3	3	24M	94.0 ms	92.01% ±1.54	84.46% ±1.73	87.32% ±0.70	84.53% ±1.12	87.32% ±2.14	92.69% ±1.13	94.62% ±0.77	51.43% ±7.93	87.75% ±0.50	90.62% ±0.59	80.00% ±4.00	80.37% ±2.99	90.71% ±1.74	78.88% ±0.96	81.56% ±1.52	78.50% ±1.05	86.97% ±2.17	91.49% ±1.27	95.08% ±0.82	50.78% ±7.75	83.88% ±1.02	90.44% ±0.54	79.78% ±4.51	80.17% ±2.88	2.82%	2.31%	5.40%	cần sửap85_not_met	–
15	e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1	intfloat/multilingual-e5-base	0.017	m1	encoder	zephyrus	1	1	86M	11.7 ms	90.71%	79.27%	87.90%	81.93%	85.46%	91.58%	93.85%	46.92%	86.25%	89.75%	81.54%	76.26%	88.98%	74.49%	83.22%	76.93%	84.82%	90.04%	93.44%	45.60%	82.94%	89.53%	81.97%	75.76%	2.31%	1.54%	6.19%	okp85_not_met	–
16	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m1	jhu-clsp/mmBERT-base	0.155	m1	encoder	zephyrus	1	1	111M	22.4 ms	92.96%	83.45%	88.36%	85.29%	87.53%	93.20%	86.15%	60.22%	85.76%	89.17%	78.46%	71.83%	91.71%	79.31%	85.34%	81.46%	87.09%	92.03%	86.07%	58.99%	82.33%	88.79%	77.87%	70.90%	3.08%	9.23%	3.32%	okp85_not_met	–
17	e2_qwen3_17b_base_genlabel_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.081	m1	decoder	legion	1	1	17M	55.4 ms	92.54%	82.85%	91.73%	86.04%	87.53%	93.13%	91.54%	51.97%	85.62%	89.42%	77.69%	70.63%	91.19%	79.38%	89.98%	82.94%	87.03%	91.94%	90.98%	50.45%	82.31%	89.25%	77.87%	69.85%	1.54%	4.62%	4.94%	okp85_not_met	–
18	e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-1.7B-Base	0.105	m1	decoder	legion	3	3	17M	52.8 ms	92.66% ±0.31	85.55% ±0.95	92.30% ±0.29	87.81% ±0.36	87.99% ±0.29	93.28% ±0.29	93.33% ±0.44	51.49% ±0.43	87.67% ±1.00	90.64% ±0.46	83.85% ±1.33	72.96% ±5.26	91.38% ±0.28	82.85% ±1.06	90.52% ±0.26	85.44% ±0.49	87.51% ±0.28	92.11% ±0.26	92.90% ±0.47	50.68% ±0.78	84.27% ±1.02	90.32% ±0.49	83.88% ±1.71	72.38% ±5.21	2.31%	3.08%	5.13%	okp85_not_met	–
19	e3_ensemble_top_greedy_m1	–	0.113	m1	ensemble	zephyrus	1	1	N/A	149.4 ms	93.30%	85.29%	92.72%	87.87%	89.15%	93.81%	94.62%	54.42%	89.50%	92.74%	85.38%	84.09%	92.17%	82.08%	91.19%	85.13%	88.82%	92.81%	95.08%	53.46%	86.55%	92.75%	86.07%	84.00%	2.31%	2.31%	4.62%	okp85_not_met	–
20	e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m1	intfloat/multilingual-e5-large-instruct	0.017	m1	encoder	legion	3	3	304M	12.3 ms	92.86% ±0.89	84.13% ±0.36	90.35% ±0.73	86.18% ±0.35	88.36% ±1.54	93.33% ±0.67	94.36% ±0.44	56.58% ±7.56	86.40% ±0.81	90.59% ±0.13	82.82% ±1.18	76.43% ±2.61	91.66% ±1.01	80.29% ±0.63	87.35% ±0.81	82.53% ±0.21	87.99% ±1.56	92.23% ±0.76	94.54% ±0.95	55.64% ±7.42	82.56% ±0.80	90.43% ±0.08	83.33% ±1.71	75.96% ±2.55	1.79%	2.56%	4.34%	okp85_not_met	–
21	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1	jhu-clsp/mmBERT-base	0.108	m1	encoder	zephyrus	3	3	111M	17.0 ms	90.57% ±1.30	82.11% ±0.53	88.67% ±2.49	83.77% ±1.90	85.01% ±1.39	91.49% ±0.90	92.56% ±5.56	46.27% ±6.80	86.32% ±0.99	88.64% ±2.33	86.41% ±6.41	68.16% ±13.03	89.39% ±1.35	78.36% ±1.26	85.55% ±4.83	80.16% ±3.05	84.82% ±1.28	90.36% ±0.95	92.08% ±5.93	46.89% ±6.38	83.11% ±1.19	88.45% ±2.27	86.61% ±6.37	68.00% ±12.26	3.08%	3.85%	6.49%	oktest_informedp85_not_met	–
22	e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1	Qwen/Qwen3-4B-Base	0.056	m1	decoder	legion	3	3	33M	138.1 ms	93.00% ±0.35	83.41% ±0.33	90.97% ±0.74	86.01% ±0.19	88.90% ±0.38	93.51% ±0.27	95.38% ±0.77	54.20% ±1.93	88.04% ±0.12	91.56% ±0.59	80.00% ±3.35	79.79% ±4.68	91.85% ±0.37	79.17% ±0.30	87.35% ±1.43	81.84% ±0.58	88.46% ±0.45	92.46% ±0.28	95.08% ±0.82	53.35% ±1.93	84.49% ±0.35	91.41% ±0.68	80.05% ±3.10	80.06% ±5.05	1.54%	1.79%	4.71%	okp85_not_met	–
23	e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1	unsloth/gemma-3-1b-pt	0.045	m1	decoder	legion	2 1 lỗi	2	13M	51.1 ms	92.15% ±0.12	84.98% ±2.72	88.75% ±1.31	85.57% ±0.56	87.30% ±0.72	92.83% ±0.09	95.00% ±0.54	50.21% ±1.15	87.49% ±0.03	90.71% ±0.79	85.00% ±2.72	75.17% ±0.24	90.83% ±0.29	81.27% ±3.94	84.37% ±1.40	81.29% ±0.91	86.93% ±0.87	91.63% ±0.23	95.49% ±0.58	49.49% ±1.64	84.08% ±0.30	90.56% ±0.83	85.66% ±2.90	74.64% ±0.26	2.69%	2.31%	5.50%	cần sửap85_not_met	–
24	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepeval_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1, nhãn m1	vinai/phobert-base-v2	0.033	m1	encoder	zephyrus	3	3	86M	19.2 ms	91.46% ±0.35	82.67% ±1.68	86.91% ±0.87	83.55% ±1.24	86.35% ±0.07	92.18% ±0.23	91.79% ±1.60	48.47% ±2.32	87.90% ±1.26	90.80% ±0.74	83.33% ±1.18	77.76% ±0.96	90.22% ±0.46	77.05% ±3.17	81.12% ±2.37	77.76% ±2.79	86.10% ±0.15	91.02% ±0.31	91.80% ±1.42	48.25% ±2.54	84.90% ±1.50	90.58% ±0.79	83.06% ±1.25	77.17% ±1.01	3.85%	3.08%	5.72%	cần sửatest_informedp85_not_met	–
25	e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1	unsloth/gemma-3-1b-pt	0.045	m1	decoder	legion	3	3	26M	49.5 ms	91.42% ±1.09	82.69% ±2.04	89.35% ±1.90	84.63% ±2.15	86.62% ±1.66	92.19% ±0.78	94.10% ±0.89	48.73% ±5.82	85.59% ±0.76	90.05% ±0.29	80.00% ±2.77	77.81% ±1.36	89.98% ±1.39	78.82% ±3.00	85.85% ±2.81	80.72% ±3.03	86.23% ±1.80	90.92% ±1.01	94.54% ±0.95	47.68% ±6.05	81.76% ±0.69	89.88% ±0.06	80.05% ±1.71	77.34% ±1.87	2.82%	2.56%	5.88%	okp85_not_met	–
26	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1	jhu-clsp/mmBERT-base	0.052	m1	encoder	zephyrus	3	3	111M	18.0 ms	91.72% ±0.71	81.07% ±0.99	90.20% ±1.28	83.89% ±0.91	86.38% ±0.83	92.34% ±0.57	91.03% ±1.60	51.25% ±3.98	85.64% ±0.46	89.96% ±0.60	81.03% ±4.89	75.90% ±4.12	90.54% ±0.71	77.82% ±0.89	88.55% ±2.51	80.76% ±1.12	86.13% ±0.81	91.25% ±0.60	90.98% ±1.42	51.33% ±3.37	82.46% ±1.09	89.74% ±0.68	81.15% ±4.99	75.32% ±4.38	2.56%	4.62%	5.10%	okp85_not_met	–
27	e2_mmbert_base_cross_ce_synth1x_frozenemb_lr2e5_fbmatched_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	jhu-clsp/mmBERT-base	0.051	m1	encoder	zephyrus	3	3	111M	18.6 ms	92.56% ±0.73	83.95% ±1.10	90.04% ±0.73	85.91% ±1.20	87.60% ±0.69	93.10% ±0.56	91.79% ±2.91	53.27% ±4.30	86.61% ±1.30	89.84% ±0.56	77.95% ±2.91	74.88% ±1.88	91.35% ±0.77	80.25% ±1.41	87.49% ±1.52	82.53% ±1.69	87.23% ±0.66	92.00% ±0.59	91.80% ±2.84	52.44% ±3.92	83.40% ±1.46	89.65% ±0.59	78.14% ±3.10	74.29% ±1.98	2.31%	4.36%	4.76%	oktest_informedp85_not_met	–
28	e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.118	m1	encoder	zephyrus	3	3	111M	18.7 ms	92.17% ±0.94	83.31% ±1.98	89.49% ±1.51	85.21% ±1.93	87.05% ±1.05	92.77% ±0.66	92.82% ±3.79	52.13% ±6.74	86.87% ±0.14	89.83% ±0.74	82.56% ±1.78	73.00% ±6.33	90.83% ±1.16	79.43% ±2.60	86.45% ±2.82	81.54% ±2.80	86.64% ±1.14	91.56% ±0.81	92.90% ±3.87	51.28% ±7.01	83.60% ±0.44	89.55% ±0.81	82.51% ±1.71	72.25% ±6.61	2.05%	4.36%	5.12%	okp85_not_met	–
29	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1	vinai/phobert-base-v2	0.008	m1	encoder	zephyrus	1	1	86M	12.0 ms	92.20%	87.54%	78.83%	77.27%	87.00%	92.59%	93.08%	53.07%	85.80%	90.02%	81.54%	80.30%	91.30%	87.60%	78.82%	77.30%	87.05%	91.76%	94.26%	54.25%	82.60%	90.03%	82.79%	80.80%	3.08%	3.08%	4.80%	okp85_not_met	–
30	e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1	Qwen/Qwen3-4B-Base	0.162	m1	decoder	kaggle	1	1	33M	233.2 ms	92.32%	82.44%	94.13%	86.48%	87.77%	92.91%	96.15%	51.44%	88.11%	90.25%	83.85%	70.78%	91.04%	79.38%	93.81%	83.88%	87.32%	91.76%	95.90%	50.43%	85.48%	89.94%	83.61%	69.86%	0.00%	1.54%	5.30%	okp85_not_met	–
31	e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1	Qwen/Qwen3-8B-Base	0.170	m1	decoder	kaggle	1	1	44M	379.1 ms	94.02%	84.49%	93.10%	87.97%	89.85%	94.30%	92.31%	61.86%	87.95%	91.35%	83.08%	77.70%	93.05%	81.08%	91.51%	85.02%	89.52%	93.40%	91.80%	60.87%	84.63%	91.04%	81.97%	76.92%	3.08%	3.08%	3.32%	okp85_not_met	–
32	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_m1	vinai/phobert-base-v2	0.073	m1	encoder	zephyrus	3	3	86M	20.4 ms	92.28% ±0.26	82.99% ±0.38	90.19% ±1.95	85.28% ±0.80	87.49% ±0.69	92.79% ±0.21	94.87% ±0.44	53.03% ±1.27	86.94% ±1.28	90.36% ±0.38	83.59% ±1.78	74.62% ±2.00	91.09% ±0.22	78.80% ±0.53	86.76% ±4.08	81.12% ±2.05	87.20% ±0.60	91.67% ±0.20	95.08% ±0.00	52.74% ±1.06	83.82% ±1.78	90.07% ±0.37	83.33% ±1.89	74.07% ±2.14	1.54%	2.05%	4.91%	oktest_informedp85_not_met	–
33	e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	decoder	legion	3	3	13M	49.8 ms	92.55% ±0.09	83.40% ±0.72	88.72% ±1.42	85.09% ±0.94	87.87% ±0.58	93.03% ±0.18	93.33% ±2.91	54.46% ±1.40	86.55% ±1.45	90.61% ±0.75	84.36% ±2.70	77.42% ±2.03	91.31% ±0.11	78.76% ±0.87	84.34% ±1.42	80.48% ±1.04	87.52% ±0.57	91.89% ±0.22	93.44% ±2.84	53.74% ±1.35	82.28% ±1.37	90.50% ±0.76	84.97% ±2.88	77.17% ±1.83	3.59%	2.82%	4.56%	okp85_not_met	–
34	e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-base	0.031	m1	encoder	zephyrus	3	3	86M	9.0 ms	91.91% ±0.94	82.49% ±0.17	86.97% ±0.50	83.55% ±0.48	86.93% ±1.44	92.52% ±0.73	95.13% ±1.78	52.23% ±5.41	84.34% ±0.58	88.80% ±1.11	81.79% ±3.11	69.27% ±6.88	90.44% ±1.10	76.05% ±1.04	80.56% ±1.07	77.04% ±0.70	86.38% ±1.50	91.16% ±0.84	94.81% ±1.89	51.03% ±5.53	79.90% ±1.02	88.60% ±1.09	83.06% ±4.04	68.55% ±6.73	2.82%	1.28%	5.16%	okp85_not_met	–
35	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.061	m1	encoder	zephyrus	3	3	111M	17.6 ms	91.70% ±0.84	83.02% ±1.19	88.39% ±0.70	84.42% ±0.92	86.73% ±1.02	92.37% ±0.62	93.59% ±1.18	50.36% ±5.38	86.10% ±1.69	89.83% ±1.00	79.49% ±3.11	75.00% ±6.06	90.41% ±0.97	78.70% ±1.54	84.43% ±1.14	80.16% ±1.15	86.38% ±1.07	91.19% ±0.73	93.72% ±1.25	49.81% ±5.00	82.59% ±2.10	89.57% ±1.12	79.51% ±2.96	74.23% ±6.65	2.31%	2.82%	5.48%	okp85_not_met	–
36	e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1	Qwen/Qwen3-4B-Base	0.131	m1	decoder	kaggle	3	3	33M	232.8 ms	93.62% ±0.85	84.89% ±2.32	91.85% ±0.94	87.26% ±0.77	89.57% ±1.09	93.99% ±0.64	94.36% ±2.70	60.20% ±9.16	88.39% ±0.73	91.66% ±0.56	82.56% ±2.91	80.69% ±0.77	92.58% ±0.98	81.71% ±3.26	89.31% ±1.78	83.97% ±1.39	89.25% ±1.10	93.04% ±0.76	93.99% ±2.88	59.47% ±9.20	85.10% ±0.75	91.54% ±0.53	82.51% ±2.63	80.52% ±0.88	2.31%	1.54%	3.80%	okp85_not_met	–
37	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.043	m1	encoder	zephyrus	3	3	86M	20.9 ms	91.93% ±0.46	84.32% ±3.06	88.06% ±1.96	84.91% ±1.23	87.18% ±0.73	92.51% ±0.36	92.05% ±2.47	51.47% ±2.39	86.78% ±0.44	89.86% ±0.18	84.36% ±1.18	70.91% ±1.04	90.61% ±0.48	80.94% ±5.91	83.16% ±3.59	79.93% ±2.18	86.81% ±0.68	91.28% ±0.38	91.80% ±2.17	50.77% ±2.33	83.06% ±0.95	89.56% ±0.20	83.88% ±0.95	70.42% ±1.00	4.62%	2.31%	5.09%	okp85_not_met	–
38	e2_gemma3_1b_pt_cross_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.053	m1	decoder	legion	3	3	13M	49.8 ms	92.25% ±0.73	83.82% ±1.48	88.29% ±0.62	84.98% ±0.95	87.35% ±0.99	92.81% ±0.64	93.85% ±0.00	52.69% ±2.87	86.92% ±1.84	90.49% ±0.62	82.05% ±1.94	79.51% ±2.85	91.09% ±0.78	79.48% ±1.80	84.03% ±0.60	80.59% ±0.98	87.14% ±0.92	91.72% ±0.70	94.54% ±0.47	52.60% ±2.64	83.43% ±2.75	90.38% ±0.70	82.51% ±2.06	79.58% ±2.85	3.08%	2.56%	4.94%	okp85_not_met	–
39	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1	vinai/phobert-base-v2	0.037	m1	encoder	zephyrus	3	3	86M	24.1 ms	92.21% ±0.47	85.37% ±1.94	90.06% ±1.50	86.57% ±0.12	87.40% ±0.45	92.72% ±0.37	92.82% ±1.78	53.11% ±2.22	88.80% ±0.77	90.97% ±0.40	82.05% ±2.47	79.20% ±2.04	90.97% ±0.59	83.02% ±3.69	87.05% ±2.98	83.38% ±0.37	87.07% ±0.55	91.56% ±0.49	92.90% ±2.37	52.59% ±2.41	86.14% ±0.75	90.71% ±0.37	81.69% ±2.63	78.88% ±1.82	2.56%	3.33%	4.79%	okp85_not_met	–
40	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m1	vinai/phobert-base-v2	0.085	m1	encoder	zephyrus	3	3	86M	20.7 ms	92.92% ±0.39	84.36% ±1.47	89.91% ±2.22	86.33% ±0.93	88.16% ±0.35	93.25% ±0.36	91.54% ±0.77	58.20% ±2.08	87.41% ±1.14	90.57% ±0.31	82.31% ±1.54	79.65% ±1.20	91.81% ±0.57	80.36% ±2.18	86.51% ±4.32	82.44% ±2.55	87.90% ±0.49	92.19% ±0.53	91.80% ±0.82	57.72% ±2.54	84.05% ±1.07	90.31% ±0.36	81.97% ±1.64	79.15% ±1.28	2.82%	4.10%	3.85%	okp85_not_met	–
41	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_m1	vinai/phobert-base-v2	0.028	m1	encoder	zephyrus	3	3	86M	22.9 ms	91.82% ±0.32	81.49% ±1.78	89.13% ±3.41	83.85% ±2.30	86.93% ±0.07	92.46% ±0.25	93.08% ±1.33	50.22% ±1.26	85.81% ±2.10	91.10% ±0.75	80.77% ±2.66	80.88% ±3.80	90.49% ±0.39	76.90% ±3.61	85.66% ±6.85	79.38% ±4.64	86.56% ±0.11	91.27% ±0.31	93.17% ±0.95	49.43% ±1.01	82.26% ±2.86	90.91% ±0.78	80.33% ±2.84	80.87% ±3.76	2.82%	2.56%	5.39%	okp85_not_met	–
42	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1	jhu-clsp/mmBERT-base	0.082	m1	encoder	zephyrus	3	3	111M	17.8 ms	92.00% ±1.04	81.32% ±1.24	88.43% ±1.31	83.60% ±1.47	86.67% ±1.68	92.55% ±0.79	91.79% ±3.87	53.54% ±7.37	85.66% ±0.58	89.51% ±1.00	81.54% ±4.62	75.03% ±7.33	90.59% ±1.26	77.00% ±1.30	84.89% ±2.16	79.32% ±1.80	86.22% ±1.74	91.27% ±0.97	91.80% ±3.76	52.68% ±7.39	82.24% ±0.93	89.32% ±1.05	82.24% ±4.94	74.70% ±7.42	2.56%	4.10%	4.80%	okp85_not_met	–
43	e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.057	m1	encoder	zephyrus	3	3	111M	19.4 ms	91.52% ±0.35	81.51% ±1.72	89.89% ±0.77	84.00% ±1.10	86.21% ±0.33	92.29% ±0.31	92.56% ±2.35	48.07% ±1.10	85.04% ±1.24	89.32% ±0.68	80.77% ±3.53	70.73% ±5.40	90.18% ±0.49	77.92% ±1.88	87.76% ±1.56	80.58% ±1.55	85.87% ±0.44	91.09% ±0.41	92.62% ±2.17	47.62% ±1.47	81.64% ±1.34	89.08% ±0.73	81.15% ±3.76	70.00% ±5.57	2.31%	3.33%	5.83%	oktest_informedp85_not_met	–
44	e2_qwen3_17b_base_cross_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.101	m1	decoder	legion	3	3	17M	48.0 ms	92.01% ±0.67	82.36% ±1.60	89.72% ±2.01	84.86% ±2.02	87.54% ±1.19	92.62% ±0.47	91.54% ±2.77	51.55% ±4.46	85.64% ±0.54	89.59% ±0.59	78.21% ±0.44	70.44% ±0.51	90.71% ±0.77	78.09% ±2.30	86.25% ±3.63	80.72% ±3.03	87.18% ±1.20	91.44% ±0.54	91.80% ±2.96	50.94% ±4.41	82.25% ±1.05	89.38% ±0.67	78.42% ±0.47	70.18% ±1.04	2.82%	3.85%	5.09%	cần sửap85_not_met	–
45	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m1	vinai/phobert-base-v2	0.052	m1	encoder	zephyrus	3	3	86M	20.9 ms	92.15% ±0.33	84.37% ±1.30	88.91% ±2.46	85.40% ±1.13	87.52% ±0.66	92.72% ±0.21	93.33% ±1.18	52.32% ±4.19	87.40% ±1.88	89.70% ±0.28	81.79% ±5.00	71.12% ±4.70	90.85% ±0.48	80.03% ±2.02	84.41% ±4.90	80.57% ±2.75	87.13% ±0.73	91.51% ±0.34	93.17% ±1.71	51.65% ±4.52	84.49% ±2.60	89.42% ±0.21	81.42% ±5.33	70.79% ±4.33	2.82%	2.05%	5.01%	okp85_not_met	–
46	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1	vinai/phobert-base-v2	0.077	m1	encoder	zephyrus	3	3	86M	21.2 ms	92.71% ±1.14	86.67% ±1.09	89.14% ±1.96	86.68% ±1.42	88.27% ±1.36	93.17% ±0.93	94.10% ±0.44	55.55% ±7.84	89.47% ±0.48	91.64% ±0.02	83.85% ±1.33	81.75% ±1.51	91.56% ±1.32	84.11% ±3.06	84.34% ±3.52	81.99% ±2.62	87.99% ±1.40	92.07% ±1.10	94.26% ±0.82	55.17% ±8.02	86.64% ±0.74	91.43% ±0.10	83.88% ±1.25	81.22% ±1.71	2.31%	2.82%	4.52%	okp85_not_met	–
47	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	vinai/phobert-base-v2	0.058	m1	encoder	zephyrus	3	3	135M	19.6 ms	92.06% ±0.39	81.79% ±1.28	89.47% ±2.50	84.45% ±1.63	87.36% ±0.73	92.56% ±0.33	92.31% ±0.77	53.92% ±2.70	85.66% ±1.04	89.42% ±0.55	84.62% ±0.77	68.50% ±1.77	90.77% ±0.52	77.42% ±2.10	86.03% ±4.74	80.17% ±2.96	87.01% ±0.86	91.38% ±0.43	92.62% ±0.00	53.26% ±3.14	82.43% ±1.27	89.14% ±0.52	84.70% ±1.25	67.87% ±1.63	3.08%	2.82%	4.62%	okp85_not_met	–
48	e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1	unsloth/gemma-3-1b-pt	0.056	m1	decoder	legion	3	3	13M	49.4 ms	91.94% ±0.63	83.12% ±1.42	88.93% ±1.43	84.71% ±0.14	86.71% ±1.17	92.58% ±0.50	93.08% ±0.77	50.62% ±3.74	85.31% ±0.90	89.84% ±0.34	78.72% ±3.64	75.92% ±3.40	90.66% ±0.69	79.25% ±1.57	85.53% ±1.99	80.82% ±0.22	86.40% ±1.11	91.41% ±0.56	93.72% ±1.71	50.05% ±3.44	81.60% ±1.20	89.69% ±0.40	79.23% ±4.21	75.63% ±3.21	2.82%	3.08%	5.34%	okp85_not_met	–
49	e2_gemma3_270m_cross_ce_real_lora_m1gemma-3-270m · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn m1	unsloth/gemma-3-270m	0.098	m1	decoder	zephyrus	3	3	4M	55.8 ms	90.77% ±0.64	79.85% ±0.12	87.69% ±2.42	82.03% ±1.08	85.10% ±1.30	91.45% ±0.56	90.77% ±1.33	48.54% ±2.53	84.13% ±0.51	88.10% ±0.44	78.21% ±1.60	71.52% ±4.02	89.27% ±0.91	75.92% ±0.79	84.64% ±3.61	78.01% ±1.89	84.77% ±1.46	90.09% ±0.78	91.26% ±1.25	48.16% ±3.18	80.28% ±0.93	87.82% ±0.71	77.87% ±2.46	70.97% ±3.85	3.59%	4.10%	5.64%	okp85_not_met	–
50	e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-large-instruct	0.029	m1	encoder	legion	3	3	304M	12.6 ms	92.79% ±0.26	84.33% ±1.13	88.33% ±1.99	85.28% ±0.52	88.08% ±0.55	93.27% ±0.24	93.08% ±1.54	54.45% ±1.29	86.04% ±0.21	90.23% ±0.77	82.82% ±0.44	73.06% ±3.88	91.69% ±0.24	79.93% ±0.59	83.60% ±3.52	80.40% ±1.37	87.82% ±0.49	92.26% ±0.24	93.72% ±1.25	53.81% ±1.56	82.30% ±0.49	90.16% ±0.72	83.88% ±1.25	72.73% ±3.79	2.31%	4.10%	4.55%	cần sửap85_not_met	–
51	e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.034	m1	encoder	zephyrus	3	3	111M	19.7 ms	91.28% ±0.83	82.27% ±2.37	91.46% ±1.02	85.14% ±1.81	86.29% ±0.68	92.09% ±0.69	94.10% ±1.18	47.86% ±2.37	86.53% ±1.31	90.27% ±0.57	80.00% ±3.53	81.56% ±4.71	90.51% ±0.35	79.53% ±2.30	90.37% ±1.83	82.83% ±2.06	86.44% ±0.27	91.31% ±0.33	93.99% ±2.06	49.72% ±0.53	83.66% ±1.16	90.14% ±0.54	79.78% ±3.41	81.80% ±4.02	2.56%	2.31%	6.00%	okp85_not_met	–
52	e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.051	m1	encoder	zephyrus	3	3	111M	18.3 ms	91.56% ±0.56	81.43% ±2.76	89.76% ±3.15	83.95% ±3.23	86.35% ±0.86	92.26% ±0.49	93.08% ±0.77	49.95% ±0.97	85.06% ±2.83	89.96% ±0.68	81.03% ±0.89	76.34% ±1.01	90.49% ±0.64	78.34% ±3.08	87.78% ±5.31	81.02% ±4.19	86.24% ±1.01	91.27% ±0.56	93.72% ±0.95	50.45% ±0.76	82.07% ±2.78	89.77% ±0.76	81.15% ±0.82	75.98% ±1.28	2.31%	3.08%	5.45%	okp85_not_met	–
53	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1	jhu-clsp/mmBERT-base	0.040	m1	encoder	zephyrus	3	3	111M	16.6 ms	91.31% ±1.17	81.50% ±1.65	87.07% ±1.38	82.79% ±1.37	85.97% ±2.14	92.11% ±0.97	94.10% ±0.89	47.82% ±4.15	86.19% ±1.65	90.57% ±0.71	82.82% ±1.94	77.99% ±3.56	89.80% ±1.27	76.48% ±1.00	82.08% ±1.26	77.64% ±0.62	85.51% ±2.05	90.77% ±1.04	93.99% ±0.47	46.88% ±3.83	82.67% ±2.03	90.35% ±0.77	82.79% ±1.64	77.47% ±3.83	1.28%	3.08%	6.04%	okp85_not_met	–
54	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1	jhu-clsp/mmBERT-base	0.057	m1	encoder	zephyrus	3	3	111M	19.6 ms	91.93% ±0.91	82.27% ±1.56	89.53% ±0.41	84.47% ±1.11	87.13% ±1.07	92.57% ±0.71	93.59% ±1.18	51.27% ±4.52	85.81% ±0.69	89.98% ±0.17	77.69% ±0.77	78.52% ±1.75	90.70% ±0.69	78.73% ±1.64	86.50% ±0.43	80.92% ±1.29	86.82% ±0.84	91.44% ±0.55	93.72% ±1.25	50.96% ±3.30	82.83% ±1.15	89.73% ±0.21	77.60% ±0.47	78.05% ±1.78	1.79%	3.08%	5.25%	okp85_not_met	–
55	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.070	m1	encoder	zephyrus	1	1	111M	21.3 ms	92.75%	82.78%	91.04%	85.72%	87.69%	93.16%	92.31%	56.60%	86.57%	89.69%	83.85%	76.22%	91.66%	79.37%	89.49%	82.69%	87.46%	92.15%	92.62%	56.50%	83.25%	89.37%	83.61%	75.56%	1.54%	3.85%	4.13%	okp85_not_met	–
56	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.047	m1	encoder	zephyrus	3	3	111M	20.2 ms	91.87% ±0.32	82.88% ±2.35	87.77% ±0.81	84.20% ±1.29	86.75% ±0.70	92.48% ±0.22	91.28% ±0.89	50.94% ±2.26	85.84% ±1.50	89.46% ±0.68	78.21% ±3.11	76.81% ±1.59	90.54% ±0.41	78.47% ±3.29	83.37% ±2.16	79.67% ±2.44	86.38% ±0.69	91.26% ±0.29	91.53% ±0.95	50.31% ±2.28	81.97% ±2.36	89.17% ±0.72	77.87% ±3.28	76.18% ±1.69	2.56%	5.13%	5.15%	cần sửap85_not_met	–
57	e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1	Qwen/Qwen3-1.7B-Base	0.095	m1	decoder	legion	3	3	17M	52.9 ms	93.03% ±0.71	86.71% ±0.63	89.92% ±2.88	87.34% ±1.54	88.53% ±1.05	93.53% ±0.59	91.28% ±1.94	54.11% ±4.01	88.01% ±0.25	90.74% ±0.25	82.05% ±2.91	73.82% ±2.24	91.76% ±0.88	83.77% ±0.87	87.29% ±3.67	84.40% ±1.87	88.07% ±1.16	92.36% ±0.74	90.98% ±1.64	52.95% ±4.17	84.87% ±0.26	90.52% ±0.16	82.24% ±2.88	73.31% ±2.29	2.31%	5.64%	4.56%	okp85_not_met	–
58	e2_xlmr_base_khonly_ce_real_frozenemb_m1XLM-R base · chỉ câu KH · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1	FacebookAI/xlm-roberta-base	0.012	m1	encoder	zephyrus	1	1	86M	7.9 ms	91.22%	81.23%	87.45%	83.09%	86.38%	91.84%	94.62%	50.62%	85.53%	89.31%	80.00%	78.20%	89.65%	74.51%	81.64%	76.70%	85.89%	90.41%	95.08%	49.57%	81.86%	89.14%	80.33%	78.40%	1.54%	1.54%	5.39%	okp85_not_met	–
59	e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.042	m1	encoder	legion	3	3	100M	55.9 ms	91.63% ±0.40	84.17% ±0.15	88.57% ±1.68	85.08% ±1.02	86.24% ±0.58	92.31% ±0.29	93.85% ±2.04	49.59% ±2.84	87.09% ±0.55	90.23% ±0.62	76.92% ±4.00	82.18% ±3.93	90.41% ±0.37	80.50% ±0.60	85.14% ±3.24	81.44% ±2.00	86.00% ±0.55	91.17% ±0.29	94.26% ±2.17	49.53% ±2.41	83.49% ±0.64	89.97% ±0.60	76.50% ±4.04	81.88% ±4.15	2.31%	3.59%	5.60%	okp85_not_met	–
60	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m1	vinai/phobert-base-v2	0.024	m1	encoder	zephyrus	3	3	86M	19.5 ms	91.70% ±0.39	84.75% ±0.74	88.31% ±2.61	85.13% ±1.21	86.89% ±0.94	92.37% ±0.39	94.10% ±1.18	49.35% ±1.00	86.90% ±0.91	90.34% ±0.62	84.87% ±0.89	73.07% ±0.76	90.34% ±0.39	80.47% ±1.21	83.41% ±4.12	80.19% ±2.50	86.50% ±0.93	91.11% ±0.41	93.99% ±0.95	48.60% ±0.81	83.64% ±1.06	90.10% ±0.62	84.97% ±1.25	72.66% ±0.74	2.82%	2.05%	5.64%	okp85_not_met	–
61	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m1	vinai/phobert-base-v2	0.031	m1	encoder	zephyrus	3	3	86M	22.1 ms	92.01% ±1.08	84.43% ±0.72	88.05% ±2.85	85.03% ±1.88	86.90% ±1.73	92.56% ±0.90	92.05% ±2.22	51.77% ±5.12	86.39% ±1.43	89.45% ±1.35	80.77% ±0.77	71.71% ±5.14	90.73% ±1.26	79.83% ±1.18	83.70% ±5.12	80.26% ±3.27	86.54% ±1.75	91.37% ±1.06	92.35% ±2.37	50.93% ±5.10	83.20% ±1.67	89.11% ±1.42	80.33% ±0.82	70.82% ±5.35	2.05%	4.62%	5.09%	okp85_not_met	–
62	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m1	vinai/phobert-base-v2	0.031	m1	encoder	zephyrus	3	3	135M	22.5 ms	92.25% ±0.75	84.70% ±0.95	88.47% ±1.85	85.50% ±1.41	87.20% ±0.94	92.77% ±0.58	94.10% ±0.44	53.44% ±4.47	86.11% ±0.62	89.52% ±0.60	83.33% ±2.35	71.92% ±3.51	90.99% ±0.94	80.44% ±1.78	84.10% ±3.34	81.04% ±2.67	86.84% ±1.04	91.60% ±0.75	94.26% ±0.82	52.69% ±4.61	82.77% ±0.78	89.20% ±0.68	83.06% ±2.50	71.23% ±3.70	2.31%	2.56%	4.83%	okp85_not_met	–
63	e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1	unsloth/gemma-3-1b-pt	0.096	m1	decoder	legion	3	3	13M	49.3 ms	92.01% ±0.98	79.31% ±2.99	85.03% ±4.92	81.15% ±3.92	87.39% ±1.19	92.56% ±0.75	90.77% ±0.00	53.06% ±6.06	84.44% ±0.84	90.11% ±0.27	82.05% ±1.94	73.46% ±2.66	90.80% ±1.26	75.58% ±2.21	81.95% ±2.07	77.58% ±2.39	87.10% ±1.28	91.45% ±0.96	90.71% ±0.47	53.09% ±6.57	80.09% ±0.92	89.94% ±0.17	82.51% ±1.25	73.55% ±2.59	5.13%	3.08%	4.77%	okp85_not_met	–
64	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1	jhu-clsp/mmBERT-base	0.061	m1	encoder	zephyrus	3	3	111M	19.5 ms	91.08% ±0.41	82.19% ±2.65	88.90% ±4.50	83.72% ±1.74	86.09% ±0.72	91.91% ±0.37	93.85% ±2.04	47.34% ±1.97	86.20% ±1.14	89.55% ±0.26	83.59% ±4.24	71.68% ±3.15	89.96% ±0.65	78.07% ±3.44	85.48% ±8.07	79.65% ±4.56	85.99% ±0.82	90.85% ±0.58	94.26% ±2.17	47.88% ±2.72	82.92% ±1.81	89.34% ±0.23	83.88% ±4.51	71.33% ±3.93	2.31%	3.59%	6.10%	oktest_informedp85_not_met	–
65	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.017	m1	encoder	zephyrus	3	3	86M	14.2 ms	92.44% ±0.27	86.49% ±2.29	85.90% ±0.38	84.72% ±0.93	87.64% ±0.56	92.92% ±0.22	92.56% ±1.78	53.52% ±1.58	85.15% ±1.51	89.93% ±0.79	83.85% ±2.04	71.61% ±3.16	91.21% ±0.35	81.68% ±6.17	78.82% ±0.47	77.00% ±0.90	87.35% ±0.63	91.73% ±0.28	93.17% ±1.71	53.00% ±1.84	80.27% ±2.03	89.75% ±0.84	84.70% ±1.89	71.00% ±3.30	2.56%	3.59%	4.70%	okp85_not_met	–
66	e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1	vinai/phobert-base-v2	0.026	m1	encoder	zephyrus	3	3	135M	9.2 ms	92.83% ±0.72	85.47% ±0.50	87.24% ±1.66	85.49% ±1.46	88.06% ±1.35	93.20% ±0.57	90.77% ±1.33	57.55% ±6.05	86.01% ±0.37	90.03% ±0.46	82.82% ±4.24	73.36% ±2.56	91.74% ±0.75	80.16% ±2.46	81.66% ±3.05	79.69% ±3.02	87.85% ±1.29	92.15% ±0.59	91.80% ±1.42	57.07% ±5.69	82.11% ±0.94	89.87% ±0.51	83.61% ±4.92	72.81% ±2.64	3.59%	3.85%	3.98%	okp85_not_met	–
67	e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.065	m1	encoder	zephyrus	3	3	87M	20.9 ms	92.08% ±0.32	82.97% ±1.66	88.83% ±1.25	84.81% ±0.97	87.17% ±0.64	92.61% ±0.24	92.56% ±1.18	52.98% ±2.81	85.54% ±0.63	89.26% ±0.39	83.33% ±1.94	69.34% ±1.60	90.80% ±0.47	78.76% ±1.97	84.84% ±2.15	80.53% ±1.37	86.78% ±0.81	91.42% ±0.33	92.62% ±0.82	52.41% ±3.47	81.74% ±0.45	88.92% ±0.37	83.06% ±2.06	68.51% ±1.48	3.08%	2.31%	4.82%	okp85_not_met	–
68	e2_xlmr_base_cross_logitadjust_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.018	m1	encoder	zephyrus	3	3	86M	10.6 ms	90.80% ±0.55	80.45% ±0.91	88.30% ±0.48	82.89% ±0.72	85.62% ±0.87	91.58% ±0.43	93.59% ±1.18	47.18% ±2.50	85.03% ±0.40	88.52% ±0.60	83.33% ±3.20	67.84% ±2.61	89.17% ±0.68	75.70% ±0.88	83.78% ±0.68	78.09% ±0.82	85.13% ±0.98	90.11% ±0.54	93.99% ±1.89	46.40% ±2.80	81.87% ±0.81	88.25% ±0.61	84.43% ±3.57	67.43% ±2.21	2.31%	3.33%	6.13%	okp85_not_met	–
69	e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-large-instruct	0.026	m1	encoder	legion	3	3	304M	13.1 ms	93.19% ±0.78	83.87% ±0.71	90.44% ±0.52	86.24% ±0.88	88.85% ±0.97	93.60% ±0.59	92.05% ±1.60	58.24% ±6.47	86.40% ±0.15	90.78% ±0.18	81.54% ±1.54	77.57% ±0.55	92.00% ±0.93	79.83% ±0.68	87.39% ±0.51	82.41% ±0.89	88.45% ±0.96	92.51% ±0.69	91.80% ±2.17	57.20% ±6.36	82.51% ±0.19	90.59% ±0.23	81.69% ±1.71	77.07% ±0.61	2.56%	4.10%	3.93%	okp85_not_met	–
70	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1	jhu-clsp/mmBERT-base	0.072	m1	encoder	zephyrus	3	3	111M	22.0 ms	92.32% ±0.43	82.68% ±2.11	87.19% ±1.46	84.02% ±1.22	86.95% ±0.43	92.78% ±0.29	89.23% ±2.66	54.26% ±3.56	85.72% ±1.58	89.18% ±1.07	78.46% ±2.66	77.06% ±1.26	91.01% ±0.47	77.93% ±2.83	82.78% ±2.43	79.33% ±2.32	86.53% ±0.43	91.57% ±0.31	89.34% ±2.84	53.28% ±3.29	81.87% ±2.40	88.89% ±1.12	78.14% ±2.63	76.46% ±1.30	3.59%	5.90%	4.43%	okp85_not_met	–
71	e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1	jhu-clsp/mmBERT-base	0.082	m1	encoder	zephyrus	3	3	111M	20.5 ms	92.17% ±0.18	82.26% ±1.20	87.90% ±1.60	83.85% ±1.17	86.97% ±0.64	92.70% ±0.23	93.08% ±1.54	53.05% ±1.75	85.71% ±1.60	89.25% ±0.94	82.05% ±2.91	72.44% ±3.53	90.83% ±0.32	77.74% ±1.94	83.96% ±2.98	79.37% ±2.12	86.58% ±0.76	91.48% ±0.35	93.44% ±1.64	52.36% ±1.90	81.93% ±2.52	88.95% ±0.99	81.97% ±2.96	71.83% ±3.77	2.31%	2.82%	4.82%	okp85_not_met	–
72	e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.067	m1	encoder	legion	1	1	105M	48.8 ms	90.75%	81.88%	89.33%	83.94%	85.91%	91.69%	93.08%	44.98%	86.13%	90.02%	81.54%	75.18%	89.29%	77.96%	86.27%	80.17%	85.54%	90.38%	93.44%	44.53%	82.06%	89.58%	80.33%	74.24%	2.31%	3.85%	6.64%	okp85_not_met	–
73	e2_visobert_cross_ce_real_frozenemb_lr4e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	uitnlp/visobert	0.037	m1	encoder	zephyrus	1	1	86M	18.8 ms	88.85%	80.36%	86.56%	81.50%	83.40%	90.24%	93.08%	39.16%	84.82%	86.97%	86.92%	57.36%	87.33%	77.02%	82.05%	77.56%	83.07%	88.88%	93.44%	39.18%	81.73%	86.65%	87.70%	56.61%	3.85%	2.31%	8.44%	okp85_not_met	–
74	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1	FacebookAI/xlm-roberta-base	0.027	m1	encoder	zephyrus	1	1	86M	7.8 ms	90.88%	81.75%	87.95%	83.69%	86.02%	91.65%	89.23%	47.35%	85.38%	88.66%	81.54%	66.67%	89.50%	77.14%	83.59%	79.12%	85.72%	90.36%	90.16%	47.21%	81.44%	88.32%	81.97%	65.79%	4.62%	4.62%	5.79%	okp85_not_met	–
75	e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m1mmBERT base · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.057	m1	encoder	zephyrus	3	3	110M	36.2 ms	91.96% ±1.00	86.71% ±1.29	86.66% ±1.40	85.18% ±0.80	87.47% ±1.33	92.62% ±0.78	93.59% ±1.94	50.41% ±5.04	86.52% ±1.43	90.33% ±1.08	77.44% ±3.11	75.63% ±5.59	90.80% ±1.00	81.80% ±1.27	80.90% ±2.62	79.32% ±1.93	87.22% ±1.27	91.52% ±0.80	93.72% ±1.71	50.15% ±4.26	82.41% ±1.82	90.10% ±1.16	77.32% ±2.88	75.01% ±5.94	2.56%	3.59%	5.46%	okp85_not_met	–
76	e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.206	m1	encoder	zephyrus	3	3	86M	18.8 ms	93.33% ±1.30	84.75% ±0.82	90.66% ±1.45	86.95% ±0.46	89.28% ±1.82	93.57% ±1.01	88.72% ±5.00	66.36% ±13.37	86.51% ±0.56	90.05% ±0.53	81.28% ±3.20	74.93% ±1.63	92.22% ±1.55	81.13% ±1.80	88.17% ±2.94	83.55% ±1.64	88.94% ±1.94	92.53% ±1.22	88.80% ±4.94	65.57% ±13.58	82.72% ±1.29	89.75% ±0.56	81.15% ±3.76	74.23% ±1.76	4.36%	5.38%	2.93%	okp85_not_met	–
77	e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.034	m1	encoder	zephyrus	3	3	87M	20.6 ms	92.76% ±0.30	84.22% ±1.81	89.03% ±2.07	85.76% ±0.70	88.15% ±0.51	93.12% ±0.21	90.51% ±1.18	57.48% ±3.58	87.11% ±1.45	90.25% ±0.47	83.85% ±1.33	75.13% ±3.42	91.50% ±0.46	79.74% ±2.40	84.55% ±4.41	80.88% ±2.15	87.70% ±0.70	91.92% ±0.36	90.71% ±1.25	56.77% ±3.90	83.79% ±1.66	89.85% ±0.48	83.61% ±1.42	74.60% ±3.54	3.08%	4.87%	3.93%	okp85_not_met	–
78	e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.052	m1	encoder	zephyrus	3	3	86M	19.4 ms	92.59% ±0.95	84.31% ±1.19	86.73% ±0.59	84.57% ±0.67	87.68% ±1.69	92.99% ±0.71	91.54% ±4.28	56.25% ±7.06	86.41% ±0.53	89.74% ±0.68	79.74% ±1.18	75.30% ±0.69	91.31% ±1.10	78.82% ±1.44	80.93% ±0.63	78.76% ±0.68	87.27% ±1.70	91.79% ±0.84	91.53% ±4.66	55.26% ±6.75	82.73% ±1.05	89.41% ±0.68	79.51% ±1.42	74.62% ±0.67	3.59%	3.59%	4.29%	oktest_informedp85_not_met	–
79	e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.033	m1	encoder	zephyrus	3	3	135M	20.4 ms	92.42% ±0.70	84.62% ±1.44	86.60% ±1.13	84.59% ±0.79	87.37% ±1.37	92.83% ±0.55	92.56% ±2.47	55.37% ±4.49	86.71% ±0.59	89.62% ±0.75	80.26% ±1.18	76.34% ±1.10	91.25% ±0.78	79.32% ±3.36	80.86% ±2.41	78.83% ±2.27	87.09% ±1.33	91.70% ±0.62	92.62% ±2.84	55.01% ±4.34	83.26% ±1.04	89.24% ±0.83	79.78% ±1.25	75.46% ±1.48	3.33%	3.08%	4.41%	oktest_informedp85_not_met	–
80	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1	jhu-clsp/mmBERT-base	0.144	m1	encoder	zephyrus	3	3	111M	18.3 ms	91.86% ±0.53	82.22% ±0.32	87.69% ±1.52	83.85% ±0.78	86.80% ±0.58	92.47% ±0.38	91.54% ±0.77	51.50% ±3.72	86.11% ±1.76	89.41% ±1.42	82.56% ±4.44	71.08% ±8.76	90.42% ±0.61	77.30% ±0.60	83.22% ±2.83	78.96% ±1.73	86.33% ±0.64	91.17% ±0.44	91.53% ±1.25	50.29% ±3.64	81.91% ±2.94	89.13% ±1.56	82.51% ±4.51	70.28% ±9.09	2.05%	4.62%	5.07%	okp85_not_met	–
81	e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.082	m1	encoder	zephyrus	3	3	111M	18.6 ms	91.86% ±0.99	81.74% ±1.94	91.71% ±0.54	84.98% ±1.20	86.76% ±1.29	92.48% ±0.71	91.03% ±4.64	51.91% ±7.22	85.25% ±0.43	89.34% ±0.52	79.74% ±2.91	73.54% ±3.34	90.56% ±1.13	78.72% ±2.40	90.96% ±1.27	82.24% ±1.92	86.38% ±1.29	91.31% ±0.82	90.98% ±4.56	51.25% ±7.00	82.10% ±0.46	89.09% ±0.52	79.78% ±3.41	72.84% ±3.40	2.82%	3.33%	5.09%	cần sửap85_not_met	–
82	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1	jhu-clsp/mmBERT-base	0.108	m1	encoder	zephyrus	3	3	111M	18.7 ms	92.83% ±0.53	83.59% ±1.70	88.90% ±0.38	85.37% ±1.18	87.93% ±0.71	93.21% ±0.35	91.79% ±2.35	57.88% ±5.44	86.89% ±1.07	90.23% ±0.36	84.36% ±3.47	78.83% ±2.50	91.64% ±0.56	79.35% ±1.89	85.53% ±1.17	81.35% ±1.56	87.62% ±0.65	92.11% ±0.36	92.08% ±2.37	57.03% ±5.23	83.31% ±0.97	90.00% ±0.34	84.43% ±3.28	78.36% ±2.68	2.82%	4.87%	3.96%	okp85_not_met	–
83	e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-0.6B-Base	0.075	m1	decoder	legion	3	3	10M	45.9 ms	91.82% ±1.26	82.16% ±1.10	90.95% ±1.85	85.05% ±1.58	87.47% ±2.03	92.51% ±0.94	93.59% ±1.60	50.42% ±6.55	86.76% ±0.85	90.91% ±0.81	81.28% ±3.55	77.60% ±2.17	90.44% ±1.54	78.56% ±1.46	88.30% ±2.67	81.55% ±1.93	87.05% ±2.14	91.26% ±1.16	93.44% ±1.64	49.73% ±6.74	83.70% ±1.06	90.74% ±0.77	81.42% ±3.31	77.49% ±2.08	1.54%	3.85%	5.49%	okp85_not_met	–
84	e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1	Qwen/Qwen3-0.6B-Base	0.077	m1	encoder	legion	3	3	97M	50.7 ms	90.25% ±1.18	83.16% ±2.08	87.92% ±1.65	84.16% ±2.32	85.86% ±1.11	91.14% ±0.85	90.26% ±3.87	45.47% ±5.03	86.61% ±1.61	89.26% ±0.08	75.13% ±1.94	73.06% ±1.43	89.19% ±1.92	79.29% ±3.86	84.02% ±3.45	80.31% ±4.23	85.90% ±1.65	90.09% ±1.42	90.44% ±4.13	47.40% ±7.53	83.31% ±2.25	88.91% ±0.15	74.86% ±2.50	72.48% ±1.87	3.08%	6.15%	6.45%	cần sửap85_not_met	–
85	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_m1	vinai/phobert-base-v2	0.093	m1	encoder	zephyrus	3	3	86M	20.4 ms	92.66% ±0.26	83.33% ±2.37	91.46% ±2.21	86.17% ±0.48	87.95% ±0.21	93.11% ±0.25	92.56% ±0.44	55.82% ±1.30	87.25% ±1.22	90.33% ±0.71	84.10% ±4.24	76.45% ±5.07	91.45% ±0.31	79.83% ±1.97	89.37% ±4.24	82.81% ±0.55	87.57% ±0.25	92.00% ±0.31	92.62% ±0.82	54.88% ±1.36	84.48% ±1.43	90.09% ±0.68	83.88% ±4.51	75.97% ±5.02	2.82%	3.08%	4.28%	okp85_not_met	–
86	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m1	vinai/phobert-base-v2	0.104	m1	encoder	zephyrus	3	3	86M	22.0 ms	92.78% ±1.16	85.85% ±1.45	88.85% ±2.34	86.35% ±1.56	88.02% ±2.13	93.14% ±0.88	91.54% ±3.08	58.93% ±8.87	88.51% ±0.88	90.56% ±0.39	85.13% ±1.60	77.70% ±3.76	91.56% ±1.34	83.19% ±3.25	84.84% ±3.96	82.19% ±2.60	87.64% ±2.16	91.97% ±1.03	90.98% ±3.28	58.33% ±8.95	85.46% ±0.94	90.30% ±0.38	84.70% ±1.25	77.61% ±3.43	3.59%	3.59%	3.90%	okp85_not_met	–
87	e2_xlmr_base_cross_ce_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.044	m1	encoder	zephyrus	3	3	86M	12.1 ms	91.22% ±1.29	81.76% ±0.94	87.01% ±2.06	83.20% ±1.98	86.03% ±1.63	91.87% ±0.99	90.26% ±3.55	50.47% ±7.97	84.45% ±1.98	88.35% ±1.27	80.51% ±5.88	68.70% ±10.78	89.80% ±1.62	75.96% ±2.82	81.80% ±4.35	77.52% ±3.97	85.71% ±1.90	90.56% ±1.27	90.98% ±3.57	49.96% ±8.42	80.26% ±3.40	88.21% ±1.52	81.15% ±5.37	68.27% ±11.22	3.59%	5.13%	5.37%	okp85_not_met	–
88	e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.019	m1	encoder	zephyrus	3	3	86M	10.9 ms	91.36% ±0.79	81.57% ±1.00	89.63% ±1.14	84.31% ±1.06	86.38% ±0.99	91.95% ±0.52	92.56% ±3.20	51.70% ±5.39	84.79% ±0.24	89.11% ±0.53	81.28% ±3.87	72.56% ±6.89	89.84% ±0.99	77.46% ±1.01	86.63% ±2.61	80.41% ±1.57	85.92% ±1.11	90.55% ±0.66	93.44% ±2.46	50.80% ±5.55	80.83% ±0.42	88.92% ±0.66	82.24% ±3.41	72.17% ±7.06	1.79%	4.62%	5.15%	okp85_not_met	–
89	e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.049	m1	encoder	zephyrus	3	3	86M	18.9 ms	91.90% ±0.24	81.58% ±1.17	88.85% ±1.52	83.91% ±0.64	87.05% ±0.36	92.47% ±0.29	93.08% ±1.33	51.80% ±0.81	85.27% ±0.53	89.72% ±0.32	83.59% ±3.11	71.55% ±1.53	90.63% ±0.29	77.00% ±0.68	84.98% ±3.01	79.40% ±1.27	86.76% ±0.38	91.30% ±0.33	93.17% ±1.25	51.38% ±1.18	81.82% ±0.59	89.45% ±0.25	83.61% ±3.57	71.07% ±1.62	3.08%	1.79%	5.06%	okp85_not_met	–
90	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.159	m1	encoder	zephyrus	3	3	86M	22.1 ms	92.20% ±1.03	82.98% ±2.71	85.85% ±5.20	83.15% ±4.91	87.48% ±1.35	92.58% ±0.79	91.03% ±1.60	56.24% ±7.99	86.66% ±0.87	89.21% ±0.24	84.87% ±3.79	69.61% ±1.54	91.06% ±1.09	79.63% ±2.13	83.85% ±2.53	80.71% ±2.56	87.22% ±1.26	91.52% ±0.81	91.53% ±1.89	55.95% ±7.37	83.22% ±1.41	89.00% ±0.19	85.52% ±4.04	69.16% ±1.46	3.85%	4.10%	4.28%	okp85_not_met	–
91	e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-4B	0.061	m1	decoder	kaggle	2	2	79M	126.5 ms	90.03% ±0.24	82.18% ±4.27	85.99% ±2.38	82.11% ±1.71	84.23% ±0.88	90.71% ±0.36	90.00% ±3.26	47.85% ±1.04	85.27% ±1.84	87.00% ±2.06	69.62% ±13.60	76.22% ±5.93	88.16% ±0.29	78.88% ±4.60	82.59% ±4.41	78.03% ±1.39	83.47% ±0.92	89.00% ±0.51	89.34% ±3.48	46.58% ±1.53	82.67% ±2.70	86.43% ±2.26	68.85% ±13.91	75.56% ±6.29	3.46%	5.00%	5.72%	okp85_not_met	–
92	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.025	m1	encoder	zephyrus	1	1	86M	11.5 ms	92.15%	79.09%	77.75%	76.27%	86.53%	92.43%	89.23%	54.98%	81.71%	88.05%	78.46%	70.83%	91.14%	76.26%	77.70%	75.77%	86.47%	91.50%	90.16%	55.84%	75.40%	87.98%	80.33%	70.50%	4.62%	5.38%	4.26%	okp85_not_met	–
93	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.152	m1	decoder	zephyrus	3	3	13M	62.5 ms	90.77% ±0.87	80.02% ±0.74	86.96% ±0.61	82.32% ±0.49	84.83% ±1.41	91.24% ±0.76	84.10% ±0.44	52.38% ±5.03	83.89% ±1.00	86.45% ±0.46	73.59% ±3.11	68.33% ±1.45	89.05% ±1.08	75.94% ±0.53	84.13% ±0.97	78.33% ±0.17	84.15% ±1.54	89.65% ±0.95	83.33% ±0.47	51.01% ±5.33	80.40% ±1.69	85.89% ±0.49	72.68% ±3.31	67.34% ±1.84	7.44%	5.38%	4.52%	okp85_not_met	–
94	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m1	vinai/phobert-base-v2	0.035	m1	encoder	zephyrus	3	3	86M	20.7 ms	92.20% ±1.00	83.17% ±0.72	89.42% ±0.83	85.20% ±0.75	87.48% ±1.08	92.73% ±0.77	93.59% ±1.60	53.58% ±5.42	87.16% ±0.19	90.25% ±0.97	82.05% ±4.24	75.15% ±4.33	90.89% ±1.16	78.83% ±0.52	85.47% ±1.67	80.89% ±0.95	87.07% ±1.14	91.53% ±0.90	93.72% ±1.89	52.68% ±5.43	84.03% ±0.15	89.93% ±1.03	81.69% ±4.51	74.42% ±4.53	2.31%	2.05%	4.82%	cần sửap85_not_met	–
95	e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-base	0.052	m1	encoder	zephyrus	1	1	86M	11.1 ms	90.88%	78.48%	78.97%	76.82%	85.51%	91.53%	90.00%	47.95%	81.34%	86.57%	77.69%	58.72%	89.55%	76.62%	80.03%	77.06%	85.20%	90.33%	90.16%	48.03%	75.94%	86.42%	79.51%	58.43%	2.31%	6.92%	5.70%	okp85_not_met	–
96	e2_visobert_cross_ce_real_frozenemb_lr2e5_m1	uitnlp/visobert	0.027	m1	encoder	zephyrus	1	1	86M	17.1 ms	89.10%	78.80%	83.67%	79.68%	83.95%	90.16%	90.00%	41.79%	82.98%	87.27%	77.69%	64.74%	87.49%	71.66%	76.48%	72.43%	83.58%	88.69%	90.98%	41.57%	77.22%	86.96%	77.87%	64.19%	5.38%	3.85%	7.32%	okp85_not_met	–
97	e2_visobert_cross_ce_real_frozenemb_lr1e5_m1	uitnlp/visobert	0.186	m1	encoder	zephyrus	1	1	86M	12.5 ms	89.10%	78.08%	83.13%	79.30%	82.85%	89.95%	86.15%	44.09%	80.22%	83.95%	82.31%	50.00%	87.33%	72.52%	76.99%	73.36%	82.31%	88.32%	86.07%	43.57%	74.45%	83.63%	82.79%	50.00%	6.92%	4.62%	6.37%	okp85_not_met	–
98	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1	FacebookAI/xlm-roberta-base	0.161	m1	encoder	zephyrus	1	1	86M	12.5 ms	90.92%	77.96%	84.35%	80.10%	85.75%	91.43%	90.00%	52.70%	80.85%	86.65%	86.92%	58.25%	89.39%	71.69%	77.24%	73.36%	85.37%	90.01%	90.98%	52.11%	74.13%	86.29%	87.70%	57.53%	3.85%	5.38%	4.71%	okp85_not_met	–
99	e1_tfidf_hybrid_qwen3emb_lr_m1TF-IDF câu KH (char 1-5, word 1-2-gram, giữ từ 1 ký tự, thêm kênh char bỏ dấu) + câu bot (char 2-4) + tương tác loại câu bot × từ câu KH · vector đóng băng Qwen3-Embedding-0.6B (cặp bot + KH, pooling last_token), trọng số dense 1 · LogReg · train thật, nhãn m1	Qwen/Qwen3-Embedding-0.6B	0.149	m1	tfidf_hybrid	zephyrus	1	1	157,386	35.1 ms	89.65%	81.88%	82.32%	81.01%	83.30%	90.26%	84.62%	46.81%	83.93%	85.90%	73.08%	64.19%	88.00%	76.03%	76.37%	75.09%	82.84%	88.67%	84.43%	46.61%	79.95%	85.27%	72.13%	62.86%	6.92%	6.15%	5.61%	okp85_not_met	–
100	e1_tfidf_svmcal_ce_synth1x_fbmatched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	TF-IDF	0.207	m1	tfidf	zephyrus	1	1	117,072	4.3 ms	91.31%	84.79%	83.64%	83.64%	85.82%	91.54%	81.54%	57.61%	85.52%	86.63%	68.46%	72.95%	89.80%	79.31%	77.74%	77.77%	85.33%	90.07%	81.15%	56.90%	81.97%	86.05%	67.21%	71.93%	9.23%	6.15%	3.50%	oktest_informedp85_not_met	–
101	e1_tfidf_svmcal_ce_synth1x_nofb_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent omi_intent_repeat, callback, relationship, negate, affirm), nhãn m1	TF-IDF	0.164	m1	tfidf	zephyrus	1	1	116,442	4.2 ms	91.14%	83.23%	83.44%	82.76%	85.40%	91.44%	80.77%	54.97%	85.52%	86.63%	66.92%	74.36%	89.60%	77.29%	77.52%	76.80%	84.89%	89.95%	80.33%	54.14%	81.97%	86.06%	65.57%	73.39%	10.00%	5.38%	3.86%	oktest_informedp85_not_met	–
102	e1_tfidf_svmcal_ce_synth1x_noqparticle_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm q_particle), nhãn m1	TF-IDF	0.280	m1	tfidf	zephyrus	1	1	127,464	4.5 ms	89.86%	82.70%	86.09%	83.27%	83.92%	90.53%	84.62%	46.81%	83.88%	85.84%	77.69%	57.06%	88.26%	79.04%	82.92%	79.82%	83.48%	89.01%	84.43%	46.61%	79.90%	85.21%	77.05%	55.62%	9.23%	3.85%	5.61%	oktest_informedp85_not_met	–
103	e1_tfidf_svmcal_ce_real_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn m1	TF-IDF	0.178	m1	tfidf	zephyrus	1	1	111,636	4.5 ms	91.14%	83.10%	84.01%	82.89%	85.56%	91.50%	83.85%	54.77%	85.38%	87.07%	70.77%	74.19%	89.60%	77.21%	78.09%	76.96%	85.08%	90.01%	83.61%	54.26%	81.61%	86.51%	69.67%	73.28%	10.00%	3.08%	4.04%	okp85_not_met	–
104	e1_tfidf_svmcal_ce_synth1x_nofiller_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm filler_only), nhãn m1	TF-IDF	0.267	m1	tfidf	zephyrus	1	1	128,538	4.2 ms	89.91%	83.63%	84.91%	82.99%	83.93%	90.58%	87.69%	46.91%	84.64%	86.18%	81.54%	58.56%	88.05%	79.47%	80.31%	78.49%	83.28%	88.87%	87.70%	45.73%	80.91%	85.66%	81.15%	57.56%	6.15%	3.85%	5.79%	oktest_informedp85_not_met	–
105	e1_tfidf_svmcal_ce_synth1x_noofftopic_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm prefix_offtopic), nhãn m1	TF-IDF	0.298	m1	tfidf	zephyrus	1	1	127,548	4.3 ms	90.03%	81.83%	86.33%	83.02%	84.23%	90.64%	84.62%	48.25%	84.15%	86.16%	80.00%	58.43%	88.57%	77.98%	83.20%	79.44%	83.90%	89.22%	84.43%	48.58%	80.19%	85.57%	79.51%	57.06%	8.46%	3.85%	5.30%	oktest_informedp85_not_met	–
106	e2_dual_qwen3emb06b_concat_ce_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.227	m1	decoder	zephyrus	3	3	2M	54.5 ms	88.18% ±0.76	83.85% ±1.93	80.61% ±2.46	80.49% ±1.06	81.14% ±1.03	88.80% ±0.61	85.90% ±1.18	45.40% ±3.30	82.63% ±1.35	83.28% ±1.07	79.49% ±0.44	59.16% ±3.09	86.17% ±0.83	81.25% ±3.69	74.50% ±3.67	74.43% ±2.50	80.44% ±1.02	86.86% ±0.65	85.52% ±0.95	44.57% ±3.21	77.54% ±2.30	82.75% ±1.19	79.78% ±0.47	58.51% ±3.32	6.41%	6.41%	6.07%	okp85_not_met	–
107	e1_tfidf_svmcal_ce_synth1x_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1, nhãn m1	TF-IDF	0.282	m1	tfidf	zephyrus	1	1	128,604	4.4 ms	89.74%	81.53%	85.87%	82.56%	83.68%	90.40%	83.85%	46.58%	84.59%	86.12%	81.54%	58.24%	88.05%	77.55%	82.68%	78.84%	83.18%	88.82%	83.61%	46.15%	80.85%	85.59%	81.15%	57.23%	9.23%	3.85%	5.61%	oktest_informedp85_not_met	–
108	e1_tfidf_svmcal_ce_synth1x_matched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ intent của train thật), nhãn m1	TF-IDF	0.190	m1	tfidf	zephyrus	1	1	112,440	4.3 ms	91.09%	83.09%	83.71%	82.76%	85.40%	91.44%	83.08%	55.10%	85.77%	86.93%	69.23%	73.77%	89.50%	77.15%	77.77%	76.78%	84.86%	89.89%	82.79%	54.30%	82.21%	86.34%	68.03%	72.81%	9.23%	3.85%	3.95%	oktest_informedp85_not_met	–
109	e1_tfidf_svmcal_ce_synth1x_nomoney_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm money), nhãn m1	TF-IDF	0.310	m1	tfidf	zephyrus	1	1	127,230	4.1 ms	90.12%	82.81%	86.66%	83.74%	84.49%	90.71%	85.38%	48.68%	83.57%	85.47%	78.46%	55.14%	88.47%	79.10%	83.49%	80.24%	83.98%	89.15%	85.25%	48.15%	79.63%	84.89%	77.87%	53.98%	8.46%	3.85%	5.25%	oktest_informedp85_not_met	–
110	e1_tfidf_svmcal_ce_synth1x_fbonly_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent nlu_fallback), nhãn m1	TF-IDF	0.356	m1	tfidf	zephyrus	1	1	121,248	4.2 ms	89.19%	81.12%	86.22%	82.32%	83.40%	90.05%	85.38%	43.87%	83.16%	84.40%	83.08%	48.87%	87.69%	77.33%	83.06%	78.78%	83.11%	88.60%	85.25%	44.64%	79.60%	84.09%	82.79%	49.51%	7.69%	3.85%	6.37%	oktest_informedp85_not_met	–
–	e2_bamibert_cross_ce_real_frozenemb_lr1e5_m1	–	–	m1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e2_bamibert_cross_ce_real_frozenemb_lr2e5_m1	–	–	m1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e2_bamibert_cross_ce_real_frozenemb_lr4e5_m1	–	–	m1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e3_fbverifier_devtop1_familygeomean_fbf1_m1	–	hard	m1	ensemble	zephyrus	1	1	N/A	271.8 ms	95.59%	87.45%	92.64%	89.62%	92.37%	95.61%	N/A	83.46%	89.62%	92.37%	85.38%	83.46%	95.01%	84.11%	91.36%	86.77%	92.36%	95.06%	N/A	83.33%	86.77%	92.36%	86.07%	83.33%	5.38%	6.15%	0.99%	ok	–
–	e3_fbverifier_devtop1_familygeomean_macrof1_m1	–	hard	m1	ensemble	zephyrus	1	1	N/A	271.8 ms	95.59%	87.45%	92.64%	89.62%	92.37%	95.61%	N/A	83.46%	89.62%	92.37%	85.38%	83.46%	95.01%	84.11%	91.36%	86.77%	92.36%	95.06%	N/A	83.33%	86.77%	92.36%	86.07%	83.33%	5.38%	6.15%	0.99%	ok	–
1	e3_ensemble_mean_all	–	0.065	v1	ensemble	zephyrus	1	1	N/A	429.8 ms	90.25%	78.74%	83.10%	79.84%	84.90%	90.90%	88.78%	49.71%	82.88%	88.03%	77.55%	73.08%	88.41%	76.50%	81.49%	77.69%	84.34%	89.25%	90.80%	47.59%	81.71%	88.05%	80.46%	72.16%	0.00%	10.20%	5.78%	okp85_not_met	–
2	e3_ensemble_greedy_all	–	0.070	v1	ensemble	zephyrus	1	1	N/A	429.8 ms	90.19%	78.54%	82.95%	79.70%	84.73%	90.82%	87.76%	49.71%	83.06%	88.24%	76.53%	75.00%	88.41%	76.33%	81.35%	77.59%	84.21%	89.22%	89.66%	47.85%	81.91%	88.30%	79.31%	74.19%	0.00%	10.20%	5.71%	okp85_not_met	–
3	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5	–	0.023	v1	encoder	zephyrus	1	1	86M	11.1 ms	90.56%	74.84%	81.79%	76.74%	84.59%	91.01%	82.65%	55.10%	78.55%	85.84%	74.49%	66.97%	88.87%	72.99%	80.16%	74.81%	84.06%	89.50%	83.91%	52.52%	77.36%	85.77%	77.01%	65.69%	0.00%	12.24%	4.33%	okp85_not_met	–
4	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1	vinai/phobert-base-v2	0.011	v1	encoder	zephyrus	1	1	86M	11.9 ms	89.94%	77.52%	78.83%	77.46%	84.95%	90.34%	82.65%	51.59%	81.28%	86.87%	73.47%	68.57%	88.33%	74.58%	76.75%	74.77%	84.73%	88.83%	86.21%	50.68%	79.95%	86.70%	75.86%	67.35%	2.04%	14.29%	4.99%	okp85_not_met	–
5	e2_xlmr_base_cross_wcesqrt_real_frozenemb	FacebookAI/xlm-roberta-base	0.021	v1	encoder	zephyrus	3	3	86M	16.1 ms	90.38% ±0.27	78.15% ±2.91	83.03% ±3.07	79.82% ±2.93	85.17% ±0.71	90.85% ±0.21	84.35% ±1.56	53.17% ±2.23	80.65% ±4.10	86.53% ±0.81	75.17% ±2.57	67.47% ±2.77	88.67% ±0.16	75.95% ±3.42	81.46% ±4.06	77.76% ±3.63	84.72% ±0.49	89.27% ±0.15	86.21% ±1.99	51.41% ±1.59	78.99% ±4.91	86.31% ±0.77	76.63% ±1.76	66.30% ±2.33	1.02%	12.93%	4.79%	okp85_not_met	–
6	e3_prior_em_mmbert_bottype	–	0.082	v1	post_hoc_transductive	zephyrus	1	1	N/A	24.8 ms	84.64%	79.03%	73.12%	71.29%	74.12%	85.77%	82.65%	60.00%	71.57%	74.45%	67.35%	78.57%	81.12%	78.17%	70.01%	67.26%	72.72%	82.56%	83.91%	58.40%	67.84%	73.41%	68.97%	77.92%	0.00%	12.24%	3.55%	oktransductivep85_not_met	–
7	e3_prior_em_mmbert_global	–	0.082	v1	post_hoc_transductive	zephyrus	1	1	N/A	24.8 ms	90.01%	87.75%	77.65%	79.66%	84.17%	90.32%	85.71%	52.83%	81.80%	86.73%	74.49%	78.49%	88.18%	87.00%	75.59%	76.36%	83.63%	88.70%	89.66%	49.68%	78.94%	86.73%	77.01%	77.01%	1.02%	13.27%	4.92%	oktransductivep85_not_met	–
8	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5	jhu-clsp/mmBERT-base	0.061	v1	encoder	zephyrus	3	3	111M	24.8 ms	90.46% ±0.59	81.26% ±0.48	87.10% ±1.88	83.07% ±0.88	84.84% ±1.02	91.01% ±0.49	85.71% ±1.02	50.93% ±3.28	84.38% ±1.15	86.92% ±0.69	74.49% ±4.08	69.46% ±2.76	88.87% ±0.78	79.76% ±0.65	86.83% ±2.27	81.88% ±1.08	84.48% ±1.09	89.59% ±0.64	88.51% ±1.15	49.18% ±3.42	83.57% ±1.21	86.89% ±0.74	77.01% ±4.60	68.30% ±2.87	1.70%	10.88%	5.34%	cần sửap85_not_met	–
9	e2_xlmr_base_cross_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	encoder	zephyrus	3	3	86M	10.8 ms	90.83% ±0.09	79.78% ±3.45	82.64% ±0.70	80.54% ±1.81	85.49% ±0.33	91.16% ±0.06	82.99% ±3.58	56.89% ±0.58	80.28% ±1.54	86.04% ±0.56	71.43% ±1.02	68.05% ±2.71	89.28% ±0.19	77.73% ±3.56	81.12% ±0.72	78.61% ±1.91	85.13% ±0.38	89.70% ±0.15	84.67% ±4.04	55.39% ±0.58	78.65% ±1.77	85.85% ±0.42	73.18% ±1.76	66.88% ±2.46	2.72%	12.24%	4.05%	cần sửap85_not_met	high
10	e2_xlmr_base_cross_logitadjust_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	encoder	zephyrus	3	3	86M	14.7 ms	90.23% ±0.68	76.57% ±1.40	84.10% ±2.50	78.73% ±1.60	85.21% ±1.29	90.84% ±0.62	85.37% ±4.12	50.92% ±2.60	79.20% ±0.46	87.02% ±0.85	76.19% ±1.56	65.34% ±1.94	88.49% ±0.73	74.71% ±1.44	82.83% ±2.84	76.87% ±1.68	84.74% ±1.19	89.28% ±0.67	87.36% ±3.04	49.04% ±2.10	77.80% ±0.56	86.83% ±0.66	78.16% ±2.30	63.79% ±1.35	1.70%	11.22%	5.30%	okp85_not_met	–
11	e2_qwen3_06b_base_cross_ce_real_lora	Qwen/Qwen3-0.6B-Base	0.098	v1	decoder	legion	1	1	10M	52.7 ms	89.94%	78.57%	82.53%	79.92%	84.99%	90.35%	84.69%	54.97%	79.16%	85.77%	71.43%	70.71%	88.26%	76.38%	81.07%	77.92%	84.62%	88.77%	87.36%	53.52%	77.26%	85.44%	72.41%	69.23%	6.12%	8.16%	4.46%	okp85_not_met	–
12	e3_ensemble_stack_all	–	0.088	v1	ensemble	zephyrus	1	1	N/A	429.8 ms	90.13%	81.06%	82.00%	80.77%	84.92%	90.64%	84.69%	51.88%	80.30%	86.95%	74.49%	70.87%	88.41%	80.56%	84.66%	81.56%	84.54%	89.06%	87.36%	50.67%	78.92%	86.71%	75.86%	69.47%	6.12%	9.18%	5.06%	okp85_not_met	–
13	e2_qwen3_17b_base_cross_ce_real_lora	Qwen/Qwen3-1.7B-Base	0.083	v1	decoder	legion	3	3	17M	53.1 ms	90.19% ±0.21	80.63% ±1.34	86.84% ±4.00	82.43% ±2.08	85.03% ±0.33	90.76% ±0.09	83.67% ±2.04	50.23% ±2.72	83.16% ±1.95	87.57% ±0.74	74.15% ±2.57	70.98% ±4.15	88.56% ±0.28	79.00% ±1.80	86.45% ±4.90	81.10% ±2.62	84.66% ±0.40	89.29% ±0.12	85.82% ±1.76	48.72% ±2.79	81.94% ±2.51	87.45% ±0.90	75.86% ±3.04	69.94% ±4.53	1.70%	13.27%	5.36%	okp85_not_met	–
14	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5	vinai/phobert-base-v2	0.022	v1	encoder	zephyrus	3	3	86M	12.6 ms	90.85% ±0.26	83.85% ±2.75	80.68% ±1.26	81.05% ±1.88	85.59% ±0.64	91.20% ±0.26	84.35% ±1.56	55.00% ±0.97	80.14% ±0.83	87.11% ±0.86	73.81% ±0.59	69.39% ±2.40	89.28% ±0.39	81.46% ±4.29	78.67% ±1.75	78.42% ±2.65	85.22% ±0.70	89.73% ±0.37	86.59% ±1.33	53.32% ±1.20	78.33% ±1.37	87.02% ±0.94	76.25% ±0.66	68.21% ±2.53	1.70%	12.59%	4.44%	cần sửap85_not_met	–
15	e2_me5_base_cross_ce_real_frozenemb_lr1e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	intfloat/multilingual-e5-base	0.088	v1	encoder	zephyrus	1	1	86M	11.3 ms	90.01%	77.11%	82.56%	78.85%	84.62%	90.48%	87.76%	53.75%	79.53%	85.44%	75.51%	66.07%	88.03%	74.76%	80.55%	76.51%	83.82%	88.65%	87.36%	51.01%	77.81%	84.80%	74.71%	63.73%	2.04%	8.16%	4.86%	okp85_not_met	–
16	e2_me5_base_cross_ce_real_frozenemb_lr4e5	–	0.028	v1	encoder	zephyrus	1	1	86M	11.6 ms	90.81%	83.13%	82.94%	82.32%	85.45%	91.17%	85.71%	55.63%	80.82%	86.08%	77.55%	67.26%	89.18%	81.15%	81.36%	80.33%	84.97%	89.65%	87.36%	53.52%	78.82%	85.69%	78.16%	65.38%	3.06%	10.20%	4.40%	okp85_not_met	–
17	e2_xlmr_base_khonly_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.076	v1	encoder	zephyrus	3	3	86M	8.5 ms	90.38% ±0.74	81.00% ±1.20	81.57% ±2.96	80.52% ±2.25	85.48% ±0.70	90.70% ±0.62	82.99% ±3.12	56.10% ±4.09	81.36% ±1.60	86.88% ±0.80	76.19% ±2.36	68.54% ±3.25	88.69% ±0.92	78.14% ±2.05	79.83% ±3.85	78.01% ±3.21	85.04% ±0.77	89.12% ±0.79	85.44% ±3.51	54.13% ±3.96	79.41% ±2.25	86.66% ±0.92	78.16% ±3.04	67.14% ±3.54	1.02%	15.65%	4.22%	okp85_not_met	–
18	e2_me5_base_cross_ce_real_frozenemb_lr2e5	intfloat/multilingual-e5-base	0.056	v1	encoder	zephyrus	3	3	86M	10.9 ms	90.71% ±0.40	79.46% ±3.51	80.85% ±2.02	79.53% ±2.66	85.36% ±0.62	91.10% ±0.34	84.35% ±2.12	55.12% ±1.61	78.41% ±1.30	85.67% ±0.80	73.47% ±4.68	65.15% ±1.59	89.08% ±0.44	76.65% ±4.12	78.71% ±2.51	76.89% ±3.16	84.88% ±0.67	89.60% ±0.37	85.82% ±2.39	52.97% ±1.67	76.39% ±1.47	85.47% ±0.77	75.48% ±4.79	63.64% ±1.78	3.40%	9.86%	4.42%	cần sửap85_not_met	–
19	e2_gemma3_270m_cross_ce_real_lora	unsloth/gemma-3-270m	0.116	v1	decoder	zephyrus	3	3	4M	54.2 ms	89.57% ±0.61	76.70% ±0.69	82.66% ±1.63	78.30% ±1.06	83.66% ±0.48	90.19% ±0.42	85.03% ±2.57	50.12% ±3.37	79.71% ±1.66	85.49% ±0.65	72.45% ±1.02	67.41% ±1.11	87.77% ±0.73	74.92% ±0.92	81.48% ±2.15	76.54% ±1.34	83.32% ±0.68	88.55% ±0.53	88.12% ±1.76	48.70% ±3.40	77.47% ±1.04	85.32% ±0.55	74.33% ±1.33	65.98% ±0.68	3.06%	11.22%	5.49%	okp85_not_met	–
20	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen	Qwen/Qwen3-Embedding-0.6B	0.114	v1	decoder	zephyrus	3	3	13M	82.5 ms	89.10% ±0.31	77.88% ±1.22	84.69% ±0.29	79.96% ±1.02	83.51% ±0.46	89.67% ±0.24	80.95% ±3.86	49.00% ±1.18	82.68% ±0.90	84.79% ±0.62	68.71% ±2.36	65.60% ±1.85	87.18% ±0.50	76.11% ±1.24	83.81% ±0.32	78.35% ±1.18	82.94% ±0.52	87.91% ±0.39	82.38% ±3.69	47.21% ±1.49	81.70% ±1.00	84.54% ±0.64	70.50% ±2.65	64.34% ±1.60	6.46%	9.86%	5.43%	okp85_not_met	–
21	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5	–	0.075	v1	encoder	zephyrus	1	1	111M	51.6 ms	90.44%	79.18%	82.40%	79.90%	84.97%	91.01%	84.69%	50.30%	79.79%	87.18%	76.53%	66.37%	89.10%	77.26%	81.25%	78.17%	84.91%	89.80%	88.51%	49.68%	78.18%	87.15%	79.31%	65.09%	2.04%	12.24%	5.38%	okp85_not_met	–
22	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5	–	0.026	v1	encoder	zephyrus	1	1	86M	12.9 ms	90.07%	86.82%	78.80%	80.02%	84.60%	90.41%	84.69%	53.90%	80.51%	85.95%	74.49%	69.52%	88.49%	86.28%	80.60%	81.30%	84.22%	88.95%	86.21%	52.45%	78.98%	85.55%	74.71%	67.71%	1.02%	13.27%	4.66%	okp85_not_met	–
23	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	FacebookAI/xlm-roberta-base	0.123	v1	encoder	zephyrus	1	1	86M	10.2 ms	90.44%	76.08%	78.27%	76.72%	84.80%	90.70%	77.55%	56.72%	77.49%	85.71%	74.49%	64.60%	88.87%	73.29%	75.93%	74.08%	84.45%	89.23%	79.31%	55.20%	75.18%	85.77%	78.16%	64.15%	2.04%	18.37%	3.81%	okp85_not_met	–
24	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5	jhu-clsp/mmBERT-base	0.120	v1	encoder	zephyrus	2	2	111M	23.9 ms	90.47% ±0.22	76.81% ±1.58	82.68% ±0.80	78.75% ±0.68	85.50% ±0.59	90.88% ±0.23	80.61% ±2.89	55.08% ±1.01	79.08% ±0.91	86.16% ±0.87	68.37% ±1.44	67.46% ±4.78	89.10% ±0.43	75.07% ±1.14	81.45% ±0.82	77.08% ±0.41	85.42% ±0.81	89.64% ±0.43	83.91% ±3.25	53.93% ±1.57	77.64% ±1.36	86.09% ±1.05	70.11% ±1.63	66.08% ±5.06	2.55%	13.78%	4.24%	cần sửap85_not_met	–
25	e2_visobert_cross_ce_real_frozenemb_lr2e5	–	0.093	v1	encoder	zephyrus	1	1	86M	11.8 ms	89.57%	75.59%	80.75%	77.25%	83.46%	90.05%	78.57%	52.03%	78.58%	85.07%	74.49%	65.18%	87.95%	73.67%	79.39%	75.40%	83.21%	88.57%	81.61%	50.71%	76.82%	84.92%	77.01%	63.81%	5.10%	15.31%	4.66%	okp85_not_met	–
26	e2_visobert_cross_ce_real_frozenemb_lr4e5	–	0.066	v1	encoder	zephyrus	1	1	86M	16.3 ms	89.27%	78.87%	84.39%	80.57%	83.35%	89.89%	80.61%	48.17%	81.64%	85.66%	74.49%	63.48%	87.49%	77.22%	83.88%	79.24%	83.09%	88.28%	82.76%	46.75%	80.17%	85.30%	75.86%	61.68%	2.04%	16.33%	5.58%	okp85_not_met	–
27	e1_tfidf_svmcal_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn v1	TF-IDF	0.123	v1	tfidf	zephyrus	1	1	119,328	5.4 ms	89.64%	82.90%	81.31%	81.05%	83.92%	90.12%	85.71%	50.91%	82.94%	86.20%	65.31%	74.42%	87.80%	80.90%	79.65%	79.03%	83.40%	88.40%	87.36%	49.35%	81.23%	86.04%	66.67%	73.42%	5.10%	7.14%	5.32%	okp85_not_met	–
28	e1_tfidf_logreg_balanced_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg cân bằng lớp · train thật, nhãn v1	TF-IDF	0.256	v1	tfidf	zephyrus	1	1	119,328	3.6 ms	87.97%	74.52%	84.17%	77.30%	82.10%	88.59%	81.63%	46.51%	78.22%	83.20%	71.43%	54.26%	85.57%	72.59%	82.91%	75.28%	81.10%	86.39%	82.76%	44.17%	76.50%	82.56%	73.56%	52.46%	5.10%	8.16%	6.04%	okp85_not_met	–
29	e1_tfidf_logreg_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật, nhãn v1	TF-IDF	0.125	v1	tfidf	zephyrus	1	1	119,328	3.6 ms	88.34%	78.74%	79.88%	78.23%	81.87%	88.99%	84.69%	46.63%	80.10%	84.12%	66.33%	64.36%	86.03%	76.31%	78.05%	75.89%	81.07%	86.88%	86.21%	44.38%	78.13%	83.76%	67.82%	62.77%	4.08%	8.16%	6.24%	okp85_not_met	–
30	e2_visobert_cross_ce_real_frozenemb_lr1e5	–	0.232	v1	encoder	zephyrus	1	1	86M	11.1 ms	89.08%	73.85%	79.41%	75.71%	82.28%	89.44%	71.43%	52.24%	75.92%	82.53%	66.33%	57.02%	87.34%	71.87%	77.75%	73.76%	81.84%	87.83%	72.41%	50.40%	73.90%	82.02%	66.67%	54.72%	8.16%	17.35%	4.20%	okp85_not_met	–
31	e1_tfidf_logreg_ce_synth1x_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật + synth ×1, nhãn v1	TF-IDF	0.234	v1	tfidf	zephyrus	1	1	135,678	3.6 ms	88.40%	77.19%	79.37%	77.44%	82.02%	88.88%	79.59%	48.75%	78.67%	83.49%	68.37%	58.77%	86.49%	74.98%	77.74%	75.39%	81.58%	87.06%	81.61%	47.97%	76.62%	83.06%	70.11%	57.01%	7.14%	10.20%	5.38%	oktest_informedp85_not_met	–
–	e2_mdeberta_cross_ce_real_frozenemb_lr1e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	–	–	v1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5	–	–	v1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e2_mdeberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	encoder	zephyrus	1	1	86M	29.1 ms	23.81%	3.97%	16.67%	6.41%	7.69%	9.16%	N/A	N/A	6.41%	7.69%	0.00%	N/A	26.63%	4.44%	16.67%	7.01%	8.41%	11.20%	N/A	N/A	7.01%	8.41%	0.00%	N/A	100.00%	0.00%	0.00%	cần sửainfeasible	–
–	e2_videberta_cross_ce_real_frozenemb_lr1e5	–	argmax	v1	encoder	zephyrus	1	1	86M	25.7 ms	23.81%	3.97%	16.67%	6.41%	7.69%	9.16%	N/A	N/A	6.41%	7.69%	0.00%	N/A	26.63%	4.44%	16.67%	7.01%	8.41%	11.20%	N/A	N/A	7.01%	8.41%	0.00%	N/A	100.00%	0.00%	0.00%	cần sửainfeasible	–
–	e2_videberta_cross_ce_real_frozenemb_lr2e5	–	argmax	v1	encoder	zephyrus	1	1	86M	26.8 ms	23.81%	3.97%	16.67%	6.41%	7.69%	9.16%	N/A	N/A	6.41%	7.69%	0.00%	N/A	26.63%	4.44%	16.67%	7.01%	8.41%	11.20%	N/A	N/A	7.01%	8.41%	0.00%	N/A	100.00%	0.00%	0.00%	cần sửainfeasible	–
–	e2_videberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e2_videberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	encoder	zephyrus	1	1	86M	31.6 ms	23.81%	3.97%	16.67%	6.41%	7.69%	9.16%	N/A	N/A	6.41%	7.69%	0.00%	N/A	26.63%	4.44%	16.67%	7.01%	8.41%	11.20%	N/A	N/A	7.01%	8.41%	0.00%	N/A	100.00%	0.00%	0.00%	cần sửainfeasible	–
1	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_lv21	jhu-clsp/mmBERT-base	0.076	v2_1	encoder	zephyrus	3	3	111M	42.4 ms	90.54% ±0.72	81.45% ±2.48	82.75% ±5.23	81.21% ±3.24	85.29% ±1.03	90.91% ±0.59	78.57% ±2.04	54.63% ±4.62	80.44% ±1.59	86.61% ±0.04	67.01% ±3.58	74.11% ±1.54	89.28% ±0.66	79.47% ±3.18	81.87% ±6.46	79.54% ±4.04	85.29% ±0.89	89.74% ±0.56	81.99% ±1.76	53.83% ±3.83	79.03% ±2.42	86.76% ±0.06	69.73% ±4.04	73.44% ±1.62	6.46%	13.61%	4.25%	okp85_not_met	–
2	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_lv21	vinai/phobert-base-v2	0.033	v2_1	encoder	zephyrus	1	1	86M	14.7 ms	90.62%	76.99%	77.15%	76.84%	84.21%	90.80%	77.55%	60.32%	76.69%	85.36%	68.37%	70.53%	88.95%	73.85%	74.75%	74.00%	83.80%	89.21%	79.31%	58.47%	74.18%	85.02%	68.97%	68.97%	7.14%	12.24%	3.28%	okp85_not_met	–

Số sau ±: độ lệch chuẩn mẫu (ddof=1) giữa các seed của cùng phương pháp, tính bằng điểm phần trăm. Ví dụ 84.10 ±0.37 = trung bình 84.10 qua các seed, các seed thường lệch nhau khoảng 0.4 điểm. Đây là độ dao động do seed, không phải khoảng tin cậy trên test: với cỡ test hiện tại, chênh vài điểm vẫn có thể là nhiễu (xem bootstrap). Ô không có ± là run 1 seed hoặc baseline. Cột Seed là số seed n của từng hàng.

Nền xanh đậm = tốt nhất cột, tính cả baseline (cao nhất; thấp nhất với clean fb→negate, clean fb→affirm, clean false fb, latency); hoà thì tô tất cả. Chữ xanh kèm ▲ = ứng viên vượt baseline nhưng chưa tốt nhất. Params không tô.

Precision theo từng intent

Cùng thứ tự và cùng cách tính trung bình ± độ lệch chuẩn qua seed; test_clean đứng đầu.

STT	Model_Name	Backbone	Threshold	Data	clean Affirm_Pre	clean Negate_Pre	clean Fallback_Pre	clean Callback_Pre	clean Relation_Pre	clean Repeat_Pre	reply-unseen Affirm_Pre	reply-unseen Negate_Pre	reply-unseen Fallback_Pre	reply-unseen Callback_Pre	reply-unseen Relation_Pre	reply-unseen Repeat_Pre
0	e0_prod_sft_hard_none_qc_v16_m1 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	m1	97.80%	89.56%	87.16%	96.08%	94.12%	70.59%	97.48%	89.82%	88.68%	96.04%	94.05%	54.55%
0	e0_prod_sft_hard_none_qc_v16 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v1	96.56%	89.28%	83.13%	94.67%	94.03%	62.50%	96.38%	90.20%	85.00%	94.59%	93.94%	57.14%
0	e0_prod_sft_hard_none_qc_v16_lv2 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2	96.56%	90.27%	77.11%	92.00%	97.01%	50.00%	96.24%	91.04%	78.75%	91.89%	98.48%	42.86%
0	e0_prod_sft_hard_none_qc_v16_lv21 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2_1	96.56%	89.28%	83.13%	94.67%	94.03%	62.50%	96.38%	90.20%	85.00%	94.59%	93.94%	57.14%
1	e3_ensemble_devtop1_familygeomean_m1	–	0.051	m1	98.97%	94.68%	53.14%	95.05%	92.94%	81.82%	98.91%	94.30%	52.40%	95.00%	92.86%	60.00%
2	e3_ensemble_devtop1_familygeomean_tempdev_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	98.97%	94.70%	54.74%	95.10%	92.94%	81.82%	98.92%	94.33%	53.81%	95.05%	92.86%	60.00%
3	e3_ensemble_devtop1_familygeomean_temphard_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	98.97%	94.70%	54.74%	95.10%	92.94%	81.82%	98.92%	94.33%	53.81%	95.05%	92.86%	60.00%
4	e3_ensemble_devtop1_familymean_m1	–	0.112	m1	98.70%	94.90%	55.75%	95.10%	94.12%	83.33%	98.56%	94.55%	54.84%	95.05%	94.05%	66.67%
5	e3_ensemble_sweep_k2_m1	–	0.110	m1	98.12%	94.69%	68.16%	91.43%	96.34%	76.92%	97.89%	94.33%	67.25%	91.35%	96.30%	57.14%
6	e3_ensemble_sweep_k3_m1	–	0.051	m1	98.57%	94.97%	59.33%	91.51%	95.12%	69.23%	98.48%	94.63%	58.21%	91.43%	95.06%	42.86%
7	e3_ensemble_sweep_k4_m1	–	0.058	m1	98.71%	94.96%	59.05%	93.27%	95.18%	69.23%	98.66%	94.61%	57.64%	93.20%	95.12%	42.86%
8	e3_ensemble_sweep_k5_m1	–	0.111	m1	98.65%	95.08%	67.20%	92.45%	95.18%	83.33%	98.59%	94.76%	66.67%	92.38%	95.12%	66.67%
9	e3_ensemble_sweep_k6_m1	–	0.087	m1	98.72%	95.23%	63.45%	92.45%	95.18%	90.91%	98.58%	94.92%	62.77%	92.38%	95.12%	80.00%
10	e3_ensemble_sweep_k7_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.087	m1	98.72%	95.23%	63.45%	92.45%	95.18%	90.91%	98.58%	94.92%	62.77%	92.38%	95.12%	80.00%
11	e3_ensemble_sweep_k8_m1	–	0.087	m1	98.72%	95.23%	63.45%	92.45%	95.18%	90.91%	98.58%	94.92%	62.77%	92.38%	95.12%	80.00%
12	e3_ensemble_top_geomean_m1	–	0.073	m1	98.84%	94.19%	58.22%	93.27%	95.29%	71.43%	98.83%	93.76%	56.80%	93.20%	95.24%	50.00%
13	e3_ensemble_top_mean_m1	–	0.073	m1	98.96%	94.64%	50.00%	95.10%	96.43%	71.43%	98.90%	94.26%	49.37%	95.05%	96.39%	50.00%
14	e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	98.29% ±0.26	94.13% ±0.99	51.43% ±7.93	92.63% ±2.65	93.88% ±0.95	76.39% ±10.49	98.02% ±0.34	93.70% ±1.11	50.78% ±7.75	92.56% ±2.67	93.80% ±0.96	44.44% ±9.62
15	e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1	intfloat/multilingual-e5-base	0.017	m1	98.44%	92.95%	46.92%	89.81%	91.25%	56.25%	98.12%	92.36%	45.60%	89.72%	91.14%	30.00%
16	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m1	jhu-clsp/mmBERT-base	0.155	m1	97.26%	93.70%	60.22%	90.29%	92.59%	66.67%	96.77%	92.99%	58.99%	90.20%	92.50%	44.44%
17	e2_qwen3_17b_base_genlabel_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.081	m1	98.75%	94.36%	51.97%	91.51%	91.76%	68.75%	98.54%	94.17%	50.45%	91.43%	91.67%	50.00%
18	e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-1.7B-Base	0.105	m1	98.68% ±0.19	95.08% ±0.42	51.49% ±0.43	92.39% ±2.64	93.07% ±0.96	82.60% ±3.49	98.42% ±0.24	94.60% ±0.51	50.68% ±0.78	92.00% ±2.21	92.96% ±0.97	68.45% ±5.15
19	e3_ensemble_top_greedy_m1	–	0.113	m1	98.90%	94.55%	54.42%	94.17%	96.39%	73.33%	98.82%	94.16%	53.46%	94.12%	96.34%	55.56%
20	e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m1	intfloat/multilingual-e5-large-instruct	0.017	m1	98.19% ±0.23	94.13% ±0.29	56.58% ±7.56	92.58% ±0.54	94.21% ±0.75	69.06% ±9.28	97.90% ±0.30	93.70% ±0.32	55.64% ±7.42	92.51% ±0.55	94.14% ±0.76	47.84% ±10.56
21	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1	jhu-clsp/mmBERT-base	0.108	m1	98.11% ±0.36	93.84% ±1.17	46.27% ±6.80	90.14% ±0.98	90.77% ±2.15	73.53% ±2.55	97.73% ±0.44	93.30% ±1.23	46.89% ±6.38	90.04% ±0.98	90.66% ±2.16	51.52% ±2.62
22	e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1	Qwen/Qwen3-4B-Base	0.056	m1	98.57% ±0.26	95.16% ±0.62	54.20% ±1.93	94.48% ±1.39	93.54% ±0.62	64.48% ±2.09	98.30% ±0.34	94.91% ±0.62	53.35% ±1.93	94.43% ±1.40	93.41% ±0.68	40.65% ±3.52
23	e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1	unsloth/gemma-3-1b-pt	0.045	m1	98.58% ±0.24	94.06% ±0.01	50.21% ±1.15	94.51% ±1.94	92.90% ±1.09	79.62% ±14.69	98.40% ±0.31	93.50% ±0.17	49.49% ±1.64	94.45% ±1.95	92.82% ±1.09	58.93% ±22.73
24	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepeval_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1, nhãn m1	vinai/phobert-base-v2	0.033	m1	98.27% ±0.30	93.63% ±0.31	48.47% ±2.32	92.12% ±2.54	90.65% ±1.29	72.88% ±8.39	98.06% ±0.34	93.07% ±0.32	48.25% ±2.54	92.04% ±2.56	90.90% ±1.51	40.00% ±17.32
25	e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1	unsloth/gemma-3-1b-pt	0.045	m1	98.38% ±0.19	93.47% ±1.00	48.73% ±5.82	95.31% ±2.41	90.19% ±2.87	70.04% ±11.55	98.20% ±0.21	93.02% ±1.00	47.68% ±6.05	95.57% ±2.04	90.41% ±2.95	48.06% ±16.17
26	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1	jhu-clsp/mmBERT-base	0.052	m1	98.05% ±0.14	93.59% ±0.31	51.25% ±3.98	90.93% ±0.79	93.05% ±1.52	59.53% ±10.40	97.71% ±0.18	93.00% ±0.49	51.33% ±3.37	90.84% ±0.79	92.97% ±1.54	41.07% ±7.78
27	e2_mmbert_base_cross_ce_synth1x_frozenemb_lr2e5_fbmatched_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	jhu-clsp/mmBERT-base	0.051	m1	98.20% ±0.14	94.27% ±0.29	53.27% ±4.30	94.33% ±1.43	93.17% ±2.11	70.48% ±3.43	97.93% ±0.14	93.78% ±0.39	52.44% ±3.92	94.27% ±1.44	93.08% ±2.13	50.00% ±5.56
28	e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.118	m1	98.29% ±0.50	93.98% ±0.06	52.13% ±6.74	93.70% ±1.47	91.69% ±0.85	70.09% ±8.92	98.09% ±0.63	93.30% ±0.06	51.28% ±7.01	93.63% ±1.48	91.59% ±0.86	48.68% ±13.31
29	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1	vinai/phobert-base-v2	0.008	m1	98.22%	93.10%	53.07%	91.35%	89.53%	100.00%	97.93%	92.77%	54.25%	91.26%	89.41%	100.00%
30	e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1	Qwen/Qwen3-4B-Base	0.162	m1	98.70%	95.25%	51.44%	93.33%	89.25%	66.67%	98.56%	94.91%	50.43%	93.27%	89.13%	50.00%
31	e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1	Qwen/Qwen3-8B-Base	0.170	m1	98.44%	95.36%	61.86%	93.27%	89.25%	68.75%	98.23%	95.07%	60.87%	93.20%	89.13%	50.00%
32	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_m1	vinai/phobert-base-v2	0.073	m1	98.07% ±0.13	95.16% ±0.44	53.03% ±1.27	92.67% ±1.29	88.54% ±5.45	70.46% ±8.32	97.80% ±0.13	94.59% ±0.50	52.74% ±1.06	92.60% ±1.30	88.41% ±5.50	46.63% ±2.96
33	e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	98.29% ±0.63	93.22% ±0.45	54.46% ±1.40	94.09% ±1.84	92.50% ±0.82	67.86% ±6.19	97.97% ±0.85	92.76% ±0.43	53.74% ±1.35	94.03% ±1.86	92.41% ±0.82	41.67% ±7.22
34	e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-base	0.031	m1	98.49% ±0.07	93.20% ±0.53	52.23% ±5.41	89.76% ±0.84	91.56% ±0.27	69.70% ±5.25	98.23% ±0.12	92.58% ±0.59	51.03% ±5.53	89.66% ±0.85	91.46% ±0.28	33.33% ±11.55
35	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.061	m1	97.98% ±0.43	93.99% ±1.09	50.36% ±5.38	94.75% ±2.92	92.22% ±0.27	68.81% ±7.84	97.68% ±0.52	93.45% ±1.06	49.81% ±5.00	94.70% ±2.95	92.12% ±0.27	44.44% ±9.62
36	e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1	Qwen/Qwen3-4B-Base	0.131	m1	98.53% ±0.13	94.56% ±0.44	60.20% ±9.16	93.09% ±3.58	93.47% ±2.23	69.47% ±18.94	98.28% ±0.11	94.17% ±0.48	59.47% ±9.20	93.02% ±3.61	93.39% ±2.25	51.90% ±24.43
37	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.043	m1	98.10% ±0.21	93.07% ±0.44	51.47% ±2.39	92.94% ±1.83	92.55% ±1.53	77.78% ±19.25	97.77% ±0.31	92.48% ±0.38	50.77% ±2.33	92.87% ±1.84	92.46% ±1.55	59.26% ±35.72
38	e2_gemma3_1b_pt_cross_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.053	m1	98.17% ±0.28	93.59% ±0.51	52.69% ±2.87	94.34% ±1.40	92.36% ±1.22	71.78% ±9.04	97.90% ±0.42	93.04% ±0.48	52.60% ±2.64	94.28% ±1.41	92.26% ±1.23	46.79% ±11.75
39	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1	vinai/phobert-base-v2	0.037	m1	98.11% ±0.08	93.21% ±1.41	53.11% ±2.22	91.41% ±2.31	92.94% ±0.14	83.42% ±14.47	97.78% ±0.02	92.67% ±1.56	52.59% ±2.41	91.33% ±2.33	92.86% ±0.15	70.90% ±25.21
40	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m1	vinai/phobert-base-v2	0.085	m1	97.82% ±0.46	94.22% ±0.42	58.20% ±2.08	90.37% ±0.48	90.63% ±4.10	74.94% ±7.54	97.49% ±0.61	93.63% ±0.34	57.72% ±2.54	90.28% ±0.49	90.83% ±3.62	52.22% ±13.47
41	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_m1	vinai/phobert-base-v2	0.028	m1	98.31% ±0.20	93.54% ±1.33	50.22% ±1.26	94.12% ±1.57	92.78% ±1.17	60.00% ±9.21	98.10% ±0.30	93.05% ±1.47	49.43% ±1.01	94.06% ±1.58	92.69% ±1.18	34.05% ±20.14
42	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1	jhu-clsp/mmBERT-base	0.082	m1	98.06% ±0.22	93.31% ±0.80	53.54% ±7.37	92.78% ±1.32	90.68% ±2.29	59.55% ±4.41	97.69% ±0.25	92.63% ±0.78	52.68% ±7.39	92.71% ±1.33	90.56% ±2.31	35.73% ±2.15
43	e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.057	m1	98.34% ±0.06	94.20% ±0.61	48.07% ±1.10	94.03% ±1.74	92.36% ±2.56	62.08% ±6.88	98.11% ±0.13	93.62% ±0.73	47.62% ±1.47	93.97% ±1.76	92.26% ±2.58	41.90% ±7.33
44	e2_qwen3_17b_base_cross_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.101	m1	98.03% ±0.35	94.04% ±0.71	51.55% ±4.46	93.53% ±2.36	92.67% ±0.47	64.35% ±4.38	97.68% ±0.46	93.58% ±0.79	50.94% ±4.41	93.46% ±2.38	92.59% ±0.48	40.28% ±8.67
45	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m1	vinai/phobert-base-v2	0.052	m1	98.33% ±0.33	93.27% ±2.35	52.32% ±4.19	93.21% ±0.88	92.32% ±2.87	76.79% ±10.67	98.04% ±0.48	92.60% ±2.45	51.65% ±4.52	93.15% ±0.88	92.53% ±2.38	52.22% ±13.47
46	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1	vinai/phobert-base-v2	0.077	m1	98.20% ±0.16	94.07% ±0.65	55.55% ±7.84	91.50% ±1.46	92.09% ±1.62	88.60% ±11.54	97.91% ±0.22	93.56% ±0.58	55.17% ±8.02	91.42% ±1.47	92.00% ±1.64	74.60% ±22.50
47	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	vinai/phobert-base-v2	0.058	m1	98.17% ±0.34	92.37% ±0.72	53.92% ±2.70	93.20% ±1.74	90.45% ±2.45	62.66% ±8.27	97.89% ±0.42	91.82% ±0.69	53.26% ±3.14	93.13% ±1.75	90.34% ±2.47	38.10% ±10.91
48	e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1	unsloth/gemma-3-1b-pt	0.056	m1	98.28% ±0.12	94.03% ±1.24	50.62% ±3.74	91.63% ±1.29	93.45% ±0.92	70.69% ±11.52	98.02% ±0.09	93.74% ±1.40	50.05% ±3.44	91.55% ±1.30	93.37% ±0.93	48.79% ±11.86
49	e2_gemma3_270m_cross_ce_real_lora_m1gemma-3-270m · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn m1	unsloth/gemma-3-270m	0.098	m1	97.42% ±0.74	93.08% ±0.91	48.54% ±2.53	90.94% ±0.48	93.14% ±1.78	55.98% ±3.83	96.86% ±1.00	92.51% ±1.22	48.16% ±3.18	91.45% ±0.48	93.05% ±1.80	33.46% ±4.95
50	e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-large-instruct	0.029	m1	98.27% ±0.22	94.16% ±0.90	54.45% ±1.29	93.56% ±1.42	93.81% ±1.15	71.74% ±8.79	98.08% ±0.29	93.80% ±0.87	53.81% ±1.56	93.50% ±1.43	93.74% ±1.16	46.67% ±5.77
51	e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.034	m1	98.50% ±0.22	93.44% ±0.47	47.86% ±2.37	92.37% ±3.52	92.70% ±1.83	68.74% ±9.43	98.28% ±0.31	92.93% ±0.52	49.72% ±0.53	92.30% ±3.55	92.61% ±1.85	51.34% ±11.37
52	e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.051	m1	98.10% ±0.04	93.55% ±1.22	49.95% ±0.97	94.89% ±1.47	91.38% ±1.33	60.73% ±15.51	97.81% ±0.09	93.12% ±1.37	50.45% ±0.76	94.84% ±1.49	91.27% ±1.35	42.54% ±17.01
53	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1	jhu-clsp/mmBERT-base	0.040	m1	98.24% ±0.40	94.20% ±0.46	47.82% ±4.15	93.74% ±2.00	92.03% ±3.10	62.95% ±8.76	97.92% ±0.46	93.61% ±0.44	46.88% ±3.83	93.68% ±2.02	92.32% ±2.48	34.49% ±7.86
54	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1	jhu-clsp/mmBERT-base	0.057	m1	98.27% ±0.26	93.39% ±0.55	51.27% ±4.52	94.98% ±1.03	93.02% ±0.83	62.70% ±11.25	97.98% ±0.35	92.72% ±0.55	50.96% ±3.30	94.93% ±1.04	92.94% ±0.84	42.86% ±12.37
55	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.070	m1	97.95%	94.35%	56.60%	90.48%	92.59%	64.71%	97.67%	93.72%	56.50%	90.38%	92.50%	45.45%
56	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.047	m1	97.98% ±0.23	93.61% ±0.40	50.94% ±2.26	92.84% ±2.15	93.04% ±0.59	68.86% ±14.42	97.65% ±0.26	93.04% ±0.50	50.31% ±2.28	92.77% ±2.17	92.95% ±0.60	44.07% ±19.57
57	e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1	Qwen/Qwen3-1.7B-Base	0.095	m1	98.41% ±0.38	94.09% ±0.60	54.11% ±4.01	96.29% ±0.54	93.79% ±2.21	83.60% ±5.16	98.06% ±0.48	93.56% ±0.55	52.95% ±4.17	96.25% ±0.54	93.72% ±2.23	68.06% ±6.36
58	e2_xlmr_base_khonly_ce_real_frozenemb_m1XLM-R base · chỉ câu KH · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1	FacebookAI/xlm-roberta-base	0.012	m1	97.72%	93.90%	50.62%	86.61%	91.86%	66.67%	97.44%	93.22%	49.57%	86.49%	91.76%	28.57%
59	e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.042	m1	97.93% ±0.16	94.42% ±0.70	49.59% ±2.84	92.84% ±2.59	91.75% ±1.77	78.46% ±3.41	97.56% ±0.21	94.01% ±0.80	49.53% ±2.41	92.77% ±2.61	91.65% ±1.80	57.50% ±6.61
60	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m1	vinai/phobert-base-v2	0.024	m1	98.28% ±0.47	93.71% ±0.16	49.35% ±1.00	92.45% ±1.49	92.80% ±1.26	81.94% ±6.21	98.00% ±0.53	93.17% ±0.16	48.60% ±0.81	92.38% ±1.51	92.72% ±1.28	57.94% ±8.36
61	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m1	vinai/phobert-base-v2	0.031	m1	97.74% ±0.32	94.57% ±0.19	51.77% ±5.12	92.64% ±1.76	92.11% ±2.82	77.78% ±3.85	97.41% ±0.41	94.19% ±0.23	50.93% ±5.10	92.57% ±1.77	92.02% ±2.85	51.85% ±3.21
62	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m1	vinai/phobert-base-v2	0.031	m1	98.15% ±0.29	93.61% ±0.38	53.44% ±4.47	91.13% ±1.22	92.31% ±1.79	79.58% ±2.47	97.90% ±0.38	93.06% ±0.50	52.69% ±4.61	91.05% ±1.23	92.22% ±1.81	55.71% ±5.15
63	e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1	unsloth/gemma-3-1b-pt	0.096	m1	98.17% ±0.14	92.59% ±0.50	53.06% ±6.06	94.08% ±2.48	92.19% ±1.31	45.76% ±16.77	97.78% ±0.17	92.08% ±0.57	53.09% ±6.57	94.02% ±2.50	92.06% ±1.39	24.44% ±7.70
64	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1	jhu-clsp/mmBERT-base	0.061	m1	98.16% ±0.61	93.85% ±0.27	47.34% ±1.97	92.93% ±3.41	92.34% ±1.64	68.50% ±15.63	97.92% ±0.76	93.24% ±0.19	47.88% ±2.72	92.86% ±3.44	92.65% ±2.19	43.89% ±21.10
65	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.017	m1	97.95% ±0.06	93.87% ±0.12	53.52% ±1.58	94.11% ±0.90	91.07% ±2.47	88.43% ±11.14	97.58% ±0.08	93.40% ±0.29	53.00% ±1.84	94.05% ±0.91	90.96% ±2.49	61.11% ±34.69
66	e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1	vinai/phobert-base-v2	0.026	m1	97.99% ±0.26	93.08% ±0.77	57.55% ±6.05	92.64% ±1.40	91.01% ±2.23	80.56% ±7.35	97.71% ±0.34	92.68% ±0.95	57.07% ±5.69	92.57% ±1.42	90.91% ±2.25	50.00% ±16.67
67	e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.065	m1	98.02% ±0.40	93.45% ±0.37	52.98% ±2.81	91.40% ±1.73	92.47% ±1.31	69.49% ±11.18	97.70% ±0.54	92.87% ±0.47	52.41% ±3.47	91.32% ±1.74	92.34% ±1.35	45.93% ±13.39
68	e2_xlmr_base_cross_logitadjust_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.018	m1	98.03% ±0.04	94.12% ±0.34	47.18% ±2.50	89.85% ±2.23	89.19% ±3.56	64.29% ±0.00	97.59% ±0.20	93.60% ±0.43	46.40% ±2.80	89.76% ±2.25	89.33% ±3.15	37.50% ±0.00
69	e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-large-instruct	0.026	m1	98.22% ±0.19	93.91% ±0.30	58.24% ±6.47	92.90% ±1.95	94.70% ±0.63	65.28% ±2.41	97.87% ±0.20	93.46% ±0.34	57.20% ±6.36	92.84% ±1.97	94.63% ±0.64	42.96% ±2.57
70	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1	jhu-clsp/mmBERT-base	0.072	m1	97.69% ±0.36	93.60% ±0.48	54.26% ±3.56	93.47% ±2.44	92.38% ±1.64	64.66% ±9.87	97.30% ±0.43	92.96% ±0.63	53.28% ±3.29	93.40% ±2.46	92.29% ±1.66	38.36% ±14.42
71	e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1	jhu-clsp/mmBERT-base	0.082	m1	97.78% ±0.76	94.11% ±0.61	53.05% ±1.75	94.90% ±0.84	92.91% ±1.92	60.83% ±6.82	97.35% ±0.93	93.52% ±0.59	52.36% ±1.90	94.85% ±0.85	92.82% ±1.94	35.56% ±11.76
72	e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.067	m1	98.05%	94.51%	44.98%	93.20%	93.90%	66.67%	97.70%	94.10%	44.53%	93.14%	93.83%	44.44%
73	e2_visobert_cross_ce_real_frozenemb_lr4e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	uitnlp/visobert	0.037	m1	98.49%	92.61%	39.16%	91.09%	91.57%	69.23%	98.29%	92.20%	39.18%	91.00%	91.46%	50.00%
74	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1	FacebookAI/xlm-roberta-base	0.027	m1	97.76%	92.96%	47.35%	91.51%	91.67%	69.23%	97.41%	92.38%	47.21%	91.43%	91.57%	42.86%
75	e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m1mmBERT base · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.057	m1	97.85% ±0.09	93.69% ±0.81	50.41% ±5.04	96.61% ±1.48	97.88% ±1.95	83.80% ±7.65	97.51% ±0.14	93.14% ±0.93	50.15% ±4.26	96.57% ±1.50	97.86% ±1.98	55.56% ±9.62
76	e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.206	m1	97.39% ±0.44	93.42% ±0.06	66.36% ±13.37	91.80% ±2.10	91.55% ±1.65	67.97% ±15.15	96.90% ±0.52	92.90% ±0.07	65.57% ±13.58	91.72% ±2.12	91.45% ±1.66	48.21% ±20.12
77	e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.034	m1	97.92% ±0.03	93.40% ±0.78	57.48% ±3.58	91.54% ±0.86	90.21% ±1.48	74.77% ±12.27	97.53% ±0.06	92.68% ±0.85	56.77% ±3.90	91.46% ±0.87	90.03% ±1.45	50.00% ±16.67
78	e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.052	m1	97.64% ±0.21	94.14% ±0.85	56.25% ±7.06	93.26% ±1.59	92.18% ±1.11	72.38% ±13.20	97.20% ±0.31	93.55% ±0.81	55.26% ±6.75	93.19% ±1.60	92.08% ±1.13	41.67% ±14.43
79	e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.033	m1	97.57% ±0.35	94.06% ±0.60	55.37% ±4.49	92.35% ±2.32	92.53% ±1.13	75.84% ±12.65	97.12% ±0.46	93.54% ±0.58	55.01% ±4.34	92.27% ±2.34	92.44% ±1.14	45.56% ±23.65
80	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1	jhu-clsp/mmBERT-base	0.144	m1	97.91% ±0.32	93.53% ±0.06	51.50% ±3.72	94.61% ±2.92	91.31% ±0.55	64.48% ±2.09	97.58% ±0.48	92.88% ±0.09	50.29% ±3.64	94.55% ±2.95	91.56% ±0.18	36.94% ±3.37
81	e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.082	m1	98.09% ±0.25	94.12% ±0.88	51.91% ±7.22	91.26% ±2.09	91.87% ±0.68	63.18% ±14.57	97.79% ±0.25	93.68% ±0.97	51.25% ±7.00	91.18% ±2.11	91.77% ±0.69	46.63% ±17.38
82	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1	jhu-clsp/mmBERT-base	0.108	m1	97.81% ±0.65	93.55% ±0.78	57.88% ±5.44	94.42% ±3.06	91.82% ±1.49	66.07% ±4.72	97.46% ±0.82	92.98% ±0.84	57.03% ±5.23	94.36% ±3.08	91.75% ±1.46	42.50% ±6.61
83	e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-0.6B-Base	0.075	m1	98.23% ±0.25	94.44% ±0.62	50.42% ±6.55	91.19% ±1.80	95.28% ±2.03	63.43% ±2.89	97.85% ±0.28	93.94% ±0.62	49.73% ±6.74	91.11% ±1.82	95.22% ±2.05	43.52% ±1.60
84	e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1	Qwen/Qwen3-0.6B-Base	0.077	m1	97.26% ±0.37	92.65% ±0.27	45.47% ±5.03	94.97% ±1.01	94.78% ±0.65	73.81% ±8.58	96.70% ±0.45	92.03% ±0.29	47.40% ±7.53	94.92% ±1.02	94.72% ±0.66	50.00% ±16.67
85	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_m1	vinai/phobert-base-v2	0.093	m1	98.42% ±0.36	93.40% ±1.78	55.82% ±1.30	90.78% ±2.84	91.39% ±2.67	70.19% ±10.98	98.22% ±0.37	92.96% ±1.85	54.88% ±1.36	90.69% ±2.86	91.29% ±2.70	50.95% ±8.61
86	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m1	vinai/phobert-base-v2	0.104	m1	97.81% ±0.55	93.05% ±0.45	58.93% ±8.87	91.52% ±1.56	92.58% ±1.33	81.20% ±17.07	97.34% ±0.67	92.37% ±0.61	58.33% ±8.95	91.44% ±1.57	92.49% ±1.34	67.20% ±29.11
87	e2_xlmr_base_cross_ce_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.044	m1	97.78% ±0.45	93.32% ±0.58	50.47% ±7.97	89.38% ±0.43	89.82% ±1.93	69.80% ±3.04	97.43% ±0.59	92.85% ±0.56	49.96% ±8.42	89.28% ±0.44	89.73% ±1.99	36.48% ±10.19
88	e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.019	m1	97.67% ±0.53	93.42% ±0.62	51.70% ±5.39	89.79% ±0.79	90.04% ±2.21	66.83% ±3.75	97.26% ±0.65	92.80% ±0.61	50.80% ±5.55	89.69% ±0.79	89.92% ±2.24	44.29% ±5.15
89	e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.049	m1	98.04% ±0.65	93.75% ±0.61	51.80% ±0.81	92.92% ±1.80	91.37% ±1.36	61.57% ±5.61	97.72% ±0.80	93.20% ±0.54	51.38% ±1.18	92.87% ±1.84	91.27% ±1.38	35.56% ±3.85
90	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.159	m1	97.52% ±0.31	92.87% ±0.85	56.24% ±7.99	91.80% ±1.07	92.40% ±2.08	67.07% ±8.80	97.07% ±0.43	92.51% ±0.68	55.95% ±7.37	91.72% ±1.08	92.31% ±2.10	48.21% ±9.94
91	e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-4B	0.061	m1	96.27% ±0.72	92.02% ±0.22	47.85% ±1.04	98.26% ±0.95	90.75% ±5.22	67.92% ±31.23	95.29% ±1.05	91.21% ±0.09	46.58% ±1.53	98.24% ±0.96	90.56% ±5.39	51.39% ±33.39
92	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.025	m1	97.51%	93.40%	54.98%	90.57%	88.10%	50.00%	97.08%	92.91%	55.84%	90.48%	87.95%	33.33%
93	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.152	m1	96.65% ±0.72	90.92% ±1.24	52.38% ±5.03	91.40% ±1.39	90.04% ±1.83	58.75% ±5.34	95.82% ±0.94	90.11% ±1.35	51.01% ±5.33	91.32% ±1.41	89.91% ±1.85	37.48% ±5.87
94	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m1	vinai/phobert-base-v2	0.035	m1	98.17% ±0.08	93.77% ±1.02	53.58% ±5.42	91.75% ±1.03	90.91% ±3.30	70.83% ±7.22	97.92% ±0.15	93.22% ±1.00	52.68% ±5.43	91.67% ±1.04	90.80% ±3.33	46.67% ±5.77
95	e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-base	0.052	m1	97.58%	93.23%	47.95%	87.16%	94.94%	50.00%	97.09%	92.69%	48.03%	87.04%	94.87%	40.00%
96	e2_visobert_cross_ce_real_frozenemb_lr2e5_m1	uitnlp/visobert	0.027	m1	97.55%	91.41%	41.79%	93.00%	90.70%	58.33%	97.13%	91.07%	41.57%	92.93%	90.59%	16.67%
97	e2_visobert_cross_ce_real_frozenemb_lr1e5_m1	uitnlp/visobert	0.186	m1	97.01%	89.68%	44.09%	92.08%	88.46%	57.14%	96.24%	88.99%	43.57%	92.00%	89.33%	25.00%
98	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1	FacebookAI/xlm-roberta-base	0.161	m1	96.79%	92.95%	52.70%	87.27%	91.36%	46.67%	96.13%	92.40%	52.11%	87.16%	91.25%	11.11%
99	e1_tfidf_hybrid_qwen3emb_lr_m1TF-IDF câu KH (char 1-5, word 1-2-gram, giữ từ 1 ký tự, thêm kênh char bỏ dấu) + câu bot (char 2-4) + tương tác loại câu bot × từ câu KH · vector đóng băng Qwen3-Embedding-0.6B (cặp bot + KH, pooling last_token), trọng số dense 1 · LogReg · train thật, nhãn m1	Qwen/Qwen3-Embedding-0.6B	0.149	m1	96.28%	89.39%	46.81%	92.63%	93.42%	72.73%	95.49%	88.19%	46.61%	92.55%	93.33%	40.00%
100	e1_tfidf_svmcal_ce_synth1x_fbmatched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	TF-IDF	0.207	m1	96.05%	89.25%	57.61%	91.18%	94.67%	80.00%	95.19%	88.12%	56.90%	91.09%	94.59%	50.00%
101	e1_tfidf_svmcal_ce_synth1x_nofb_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent omi_intent_repeat, callback, relationship, negate, affirm), nhãn m1	TF-IDF	0.164	m1	96.24%	89.57%	54.97%	91.18%	94.67%	72.73%	95.43%	88.46%	54.14%	91.09%	94.59%	40.00%
102	e1_tfidf_svmcal_ce_synth1x_noqparticle_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm q_particle), nhãn m1	TF-IDF	0.280	m1	96.70%	89.56%	46.81%	90.29%	95.89%	76.92%	96.03%	88.43%	46.61%	90.20%	95.83%	57.14%
103	e1_tfidf_svmcal_ce_real_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn m1	TF-IDF	0.178	m1	97.00%	88.25%	54.77%	91.18%	94.67%	72.73%	96.34%	87.00%	54.26%	91.09%	94.59%	40.00%
104	e1_tfidf_svmcal_ce_synth1x_nofiller_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm filler_only), nhãn m1	TF-IDF	0.267	m1	96.71%	90.21%	46.91%	90.29%	95.83%	81.82%	96.03%	89.13%	45.73%	90.20%	95.77%	60.00%
105	e1_tfidf_svmcal_ce_synth1x_noofftopic_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm prefix_offtopic), nhãn m1	TF-IDF	0.298	m1	96.57%	89.77%	48.25%	90.29%	94.67%	71.43%	95.87%	88.66%	48.58%	90.20%	94.59%	50.00%
106	e2_dual_qwen3emb06b_concat_ce_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.227	m1	94.67% ±0.87	89.29% ±0.73	45.40% ±3.30	95.08% ±0.46	85.73% ±1.88	92.96% ±6.12	93.36% ±1.07	88.43% ±0.69	44.57% ±3.21	95.03% ±0.46	85.53% ±1.90	80.56% ±17.35
107	e1_tfidf_svmcal_ce_synth1x_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1, nhãn m1	TF-IDF	0.282	m1	96.57%	89.68%	46.58%	90.29%	94.59%	71.43%	95.86%	88.55%	46.15%	90.20%	94.52%	50.00%
108	e1_tfidf_svmcal_ce_synth1x_matched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ intent của train thật), nhãn m1	TF-IDF	0.190	m1	96.88%	88.07%	55.10%	91.18%	94.59%	72.73%	96.17%	86.81%	54.30%	91.09%	94.52%	40.00%
109	e1_tfidf_svmcal_ce_synth1x_nomoney_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm money), nhãn m1	TF-IDF	0.310	m1	96.64%	89.60%	48.68%	90.29%	94.74%	76.92%	95.94%	88.48%	48.15%	90.20%	94.67%	57.14%
110	e1_tfidf_svmcal_ce_synth1x_fbonly_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent nlu_fallback), nhãn m1	TF-IDF	0.356	m1	96.86%	89.58%	43.87%	90.29%	94.67%	71.43%	96.16%	88.43%	44.64%	90.20%	94.59%	50.00%
–	e2_bamibert_cross_ce_real_frozenemb_lr1e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_bamibert_cross_ce_real_frozenemb_lr2e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_bamibert_cross_ce_real_frozenemb_lr4e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e3_fbverifier_devtop1_familygeomean_fbf1_m1	–	hard	m1	98.01%	93.91%	83.46%	91.59%	93.02%	64.71%	97.75%	93.68%	83.33%	91.51%	92.94%	45.45%
–	e3_fbverifier_devtop1_familygeomean_macrof1_m1	–	hard	m1	98.01%	93.91%	83.46%	91.59%	93.02%	64.71%	97.75%	93.68%	83.33%	91.51%	92.94%	45.45%
1	e3_ensemble_mean_all	–	0.065	v1	96.56%	93.59%	49.71%	93.06%	89.55%	50.00%	95.95%	93.09%	47.59%	92.96%	89.39%	40.00%
2	e3_ensemble_greedy_all	–	0.070	v1	96.35%	93.88%	49.71%	91.89%	89.39%	50.00%	95.67%	93.42%	47.85%	91.78%	89.23%	40.00%
3	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5	–	0.023	v1	95.69%	93.97%	55.10%	89.33%	87.69%	27.27%	94.91%	93.85%	52.52%	89.19%	87.50%	20.00%
4	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1	vinai/phobert-base-v2	0.011	v1	94.75%	93.31%	51.59%	95.77%	89.71%	40.00%	93.78%	92.78%	50.68%	95.71%	89.55%	25.00%
5	e2_xlmr_base_cross_wcesqrt_real_frozenemb	FacebookAI/xlm-roberta-base	0.021	v1	95.95% ±0.17	92.63% ±0.59	53.17% ±2.23	92.43% ±3.12	87.13% ±2.41	47.62% ±17.17	95.16% ±0.40	92.05% ±0.68	51.41% ±1.59	92.32% ±3.16	86.95% ±2.44	37.78% ±20.37
6	e3_prior_em_mmbert_bottype	–	0.082	v1	94.94%	95.49%	60.00%	94.03%	29.73%	100.00%	94.40%	94.15%	58.40%	93.94%	28.12%	100.00%
7	e3_prior_em_mmbert_global	–	0.082	v1	93.84%	95.36%	52.83%	94.29%	90.16%	100.00%	93.00%	95.12%	49.68%	94.20%	90.00%	100.00%
8	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5	jhu-clsp/mmBERT-base	0.061	v1	96.18% ±0.30	93.86% ±0.92	50.93% ±3.28	93.05% ±3.40	88.45% ±2.44	65.08% ±7.27	95.71% ±0.41	93.52% ±0.88	49.18% ±3.42	92.95% ±3.44	88.28% ±2.47	58.89% ±8.39
9	e2_xlmr_base_cross_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	95.85% ±0.40	91.77% ±0.67	56.89% ±0.58	92.70% ±0.70	87.31% ±1.68	54.17% ±19.09	95.07% ±0.38	91.11% ±0.74	55.39% ±0.58	92.60% ±0.71	87.12% ±1.70	45.08% ±19.55
10	e2_xlmr_base_cross_logitadjust_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	96.38% ±0.37	92.66% ±0.62	50.92% ±2.60	93.59% ±2.06	90.11% ±2.19	35.73% ±2.15	95.80% ±0.38	92.08% ±0.69	49.04% ±2.10	93.50% ±2.09	89.96% ±2.23	27.86% ±2.58
11	e2_qwen3_06b_base_cross_ce_real_lora	Qwen/Qwen3-0.6B-Base	0.098	v1	95.83%	89.32%	54.97%	94.37%	86.96%	50.00%	95.09%	88.62%	53.52%	94.29%	86.76%	40.00%
12	e3_ensemble_stack_all	–	0.088	v1	95.82%	91.06%	51.88%	97.01%	90.62%	60.00%	94.92%	90.30%	50.67%	96.97%	90.48%	60.00%
13	e2_qwen3_17b_base_cross_ce_real_lora	Qwen/Qwen3-1.7B-Base	0.083	v1	95.93% ±0.10	92.55% ±1.11	50.23% ±2.72	95.71% ±1.35	91.01% ±1.34	58.33% ±7.22	95.19% ±0.07	92.16% ±1.22	48.72% ±2.79	95.65% ±1.36	90.88% ±1.36	51.43% ±9.90
14	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5	vinai/phobert-base-v2	0.022	v1	95.73% ±0.28	92.72% ±0.41	55.00% ±0.97	94.07% ±3.23	87.83% ±2.94	77.78% ±19.25	94.91% ±0.36	92.24% ±0.48	53.32% ±1.20	93.99% ±3.28	87.65% ±2.98	66.67% ±28.87
15	e2_me5_base_cross_ce_real_frozenemb_lr1e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	intfloat/multilingual-e5-base	0.088	v1	95.93%	91.55%	53.75%	89.19%	89.39%	42.86%	94.82%	91.11%	51.01%	89.04%	89.23%	33.33%
16	e2_me5_base_cross_ce_real_frozenemb_lr4e5	–	0.028	v1	95.88%	92.37%	55.63%	93.15%	86.76%	75.00%	95.04%	92.04%	53.52%	93.06%	86.57%	66.67%
17	e2_xlmr_base_khonly_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.076	v1	95.11% ±0.18	91.66% ±1.25	56.10% ±4.09	91.48% ±0.71	90.53% ±1.47	61.11% ±9.62	94.24% ±0.36	90.97% ±1.42	54.13% ±3.96	91.37% ±0.72	90.38% ±1.49	47.78% ±13.47
18	e2_me5_base_cross_ce_real_frozenemb_lr2e5	intfloat/multilingual-e5-base	0.056	v1	96.41% ±0.24	91.29% ±0.88	55.12% ±1.61	90.28% ±1.40	88.66% ±2.78	55.00% ±18.03	95.85% ±0.33	90.77% ±0.87	52.97% ±1.67	90.15% ±1.42	88.49% ±2.82	41.67% ±22.05
19	e2_gemma3_270m_cross_ce_real_lora	unsloth/gemma-3-270m	0.116	v1	95.90% ±0.50	92.19% ±0.83	50.12% ±3.37	92.12% ±1.98	87.27% ±1.44	42.59% ±8.49	95.15% ±0.41	91.53% ±0.78	48.70% ±3.40	92.45% ±1.99	87.53% ±1.09	34.17% ±8.04
20	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen	Qwen/Qwen3-Embedding-0.6B	0.114	v1	95.24% ±0.16	91.30% ±0.23	49.00% ±1.18	89.20% ±2.94	91.10% ±2.48	51.43% ±9.90	94.20% ±0.20	90.81% ±0.25	47.21% ±1.49	89.05% ±2.97	90.97% ±2.52	44.44% ±9.62
21	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5	–	0.075	v1	96.17%	93.37%	50.30%	95.65%	89.55%	50.00%	95.72%	93.18%	49.68%	95.59%	89.39%	40.00%
22	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5	–	0.026	v1	95.01%	92.15%	53.90%	91.78%	88.06%	100.00%	94.10%	91.56%	52.45%	91.67%	87.88%	100.00%
23	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	FacebookAI/xlm-roberta-base	0.123	v1	94.69%	92.78%	56.72%	88.61%	90.32%	33.33%	93.68%	92.24%	55.20%	88.46%	90.16%	20.00%
24	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5	jhu-clsp/mmBERT-base	0.120	v1	95.81% ±0.05	91.40% ±0.84	55.08% ±1.01	95.75% ±1.96	86.36% ±4.29	36.43% ±9.09	95.35% ±0.09	90.98% ±0.94	53.93% ±1.57	95.69% ±1.99	86.72% ±3.55	27.78% ±7.86
25	e2_visobert_cross_ce_real_frozenemb_lr2e5	–	0.093	v1	95.63%	90.14%	52.03%	94.20%	84.06%	37.50%	94.96%	89.85%	50.71%	94.12%	83.82%	28.57%
26	e2_visobert_cross_ce_real_frozenemb_lr4e5	–	0.066	v1	95.51%	91.55%	48.17%	94.29%	86.57%	57.14%	94.65%	91.14%	46.75%	94.20%	86.57%	50.00%
27	e1_tfidf_svmcal_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn v1	TF-IDF	0.123	v1	95.37%	91.09%	50.91%	91.67%	93.33%	75.00%	94.34%	90.29%	49.35%	91.55%	93.22%	66.67%
28	e1_tfidf_logreg_balanced_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg cân bằng lớp · train thật, nhãn v1	TF-IDF	0.256	v1	94.66%	91.82%	46.51%	87.84%	86.30%	40.00%	93.35%	91.07%	44.17%	87.67%	85.92%	33.33%
29	e1_tfidf_logreg_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật, nhãn v1	TF-IDF	0.125	v1	95.14%	90.53%	46.63%	91.43%	88.71%	60.00%	94.01%	89.63%	44.38%	91.30%	88.52%	50.00%
30	e2_visobert_cross_ce_real_frozenemb_lr1e5	–	0.232	v1	94.77%	89.75%	52.24%	89.19%	83.82%	33.33%	93.79%	89.44%	50.40%	89.04%	83.58%	25.00%
31	e1_tfidf_logreg_ce_synth1x_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật + synth ×1, nhãn v1	TF-IDF	0.234	v1	94.39%	89.77%	48.75%	91.55%	88.71%	50.00%	93.16%	88.78%	47.97%	91.43%	88.52%	40.00%
–	e2_mdeberta_cross_ce_real_frozenemb_lr1e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	N/A	23.81%	N/A	N/A	N/A	N/A	N/A	26.63%	N/A	N/A	N/A	N/A
–	e2_videberta_cross_ce_real_frozenemb_lr1e5	–	argmax	v1	N/A	23.81%	N/A	N/A	N/A	N/A	N/A	26.63%	N/A	N/A	N/A	N/A
–	e2_videberta_cross_ce_real_frozenemb_lr2e5	–	argmax	v1	N/A	23.81%	N/A	N/A	N/A	N/A	N/A	26.63%	N/A	N/A	N/A	N/A
–	e2_videberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_videberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	N/A	23.81%	N/A	N/A	N/A	N/A	N/A	26.63%	N/A	N/A	N/A	N/A
1	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_lv21	jhu-clsp/mmBERT-base	0.076	v2_1	95.67% ±0.41	90.20% ±0.68	54.63% ±4.62	97.52% ±1.67	90.50% ±2.30	60.19% ±13.13	95.03% ±0.57	90.12% ±0.62	53.83% ±3.83	97.48% ±1.70	90.35% ±2.34	50.00% ±16.67
2	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_lv21	vinai/phobert-base-v2	0.033	v2_1	95.50%	89.63%	60.32%	94.12%	82.35%	40.00%	94.51%	88.99%	58.47%	94.03%	82.09%	25.00%

Recall theo từng intent

STT	Model_Name	Backbone	Threshold	Data	clean Affirm_Rec	clean Negate_Rec	clean Fallback_Rec	clean Callback_Rec	clean Relation_Rec	clean Repeat_Rec	reply-unseen Affirm_Rec	reply-unseen Negate_Rec	reply-unseen Fallback_Rec	reply-unseen Callback_Rec	reply-unseen Relation_Rec	reply-unseen Repeat_Rec
0	e0_prod_sft_hard_none_qc_v16_m1 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	m1	97.15%	94.22%	73.08%	100.00%	88.89%	100.00%	96.55%	94.03%	77.05%	100.00%	89.77%	100.00%
0	e0_prod_sft_hard_none_qc_v16 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v1	96.75%	92.75%	70.41%	95.95%	86.30%	100.00%	95.98%	92.80%	78.16%	97.22%	87.32%	100.00%
0	e0_prod_sft_hard_none_qc_v16_lv2 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2	96.56%	91.41%	70.33%	98.57%	89.04%	100.00%	95.84%	91.55%	76.83%	98.55%	90.28%	100.00%
0	e0_prod_sft_hard_none_qc_v16_lv21 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2_1	96.75%	92.75%	70.41%	95.95%	86.30%	100.00%	95.98%	92.80%	78.16%	97.22%	87.32%	100.00%
1	e3_ensemble_devtop1_familygeomean_m1	–	0.051	m1	95.23%	85.74%	97.69%	97.96%	87.78%	75.00%	94.05%	84.65%	98.36%	97.94%	88.64%	50.00%
2	e3_ensemble_devtop1_familygeomean_tempdev_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	95.49%	86.13%	97.69%	98.98%	87.78%	75.00%	94.31%	85.07%	98.36%	98.97%	88.64%	50.00%
3	e3_ensemble_devtop1_familygeomean_temphard_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	95.49%	86.13%	97.69%	98.98%	87.78%	75.00%	94.31%	85.07%	98.36%	98.97%	88.64%	50.00%
4	e3_ensemble_devtop1_familymean_m1	–	0.112	m1	95.63%	86.13%	96.92%	98.98%	88.89%	83.33%	94.48%	85.07%	97.54%	98.97%	89.77%	66.67%
5	e3_ensemble_sweep_k2_m1	–	0.110	m1	96.82%	89.40%	93.85%	97.96%	87.78%	83.33%	96.03%	88.70%	94.26%	97.94%	88.64%	66.67%
6	e3_ensemble_sweep_k3_m1	–	0.051	m1	96.09%	87.28%	95.38%	98.98%	86.67%	75.00%	95.09%	86.35%	95.90%	98.97%	87.50%	50.00%
7	e3_ensemble_sweep_k4_m1	–	0.058	m1	96.29%	87.09%	95.38%	98.98%	87.78%	75.00%	95.26%	86.14%	95.90%	98.97%	88.64%	50.00%
8	e3_ensemble_sweep_k5_m1	–	0.111	m1	96.95%	89.40%	96.15%	100.00%	87.78%	83.33%	96.29%	88.70%	96.72%	100.00%	88.64%	66.67%
9	e3_ensemble_sweep_k6_m1	–	0.087	m1	96.75%	88.44%	96.15%	100.00%	87.78%	83.33%	95.95%	87.63%	96.72%	100.00%	88.64%	66.67%
10	e3_ensemble_sweep_k7_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.087	m1	96.75%	88.44%	96.15%	100.00%	87.78%	83.33%	95.95%	87.63%	96.72%	100.00%	88.64%	66.67%
11	e3_ensemble_sweep_k8_m1	–	0.087	m1	96.75%	88.44%	96.15%	100.00%	87.78%	83.33%	95.95%	87.63%	96.72%	100.00%	88.64%	66.67%
12	e3_ensemble_top_geomean_m1	–	0.073	m1	95.63%	87.48%	95.38%	98.98%	90.00%	83.33%	94.40%	86.57%	95.90%	98.97%	90.91%	66.67%
13	e3_ensemble_top_mean_m1	–	0.073	m1	94.57%	84.97%	96.15%	98.98%	90.00%	83.33%	93.19%	84.01%	96.72%	98.97%	90.91%	66.67%
14	e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	93.93% ±1.55	86.64% ±1.62	94.62% ±0.77	97.28% ±2.12	84.81% ±4.21	66.67% ±8.33	92.44% ±1.82	85.64% ±1.79	95.08% ±0.82	97.25% ±2.15	85.61% ±4.30	33.33% ±16.67
15	e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1	intfloat/multilingual-e5-base	0.017	m1	92.11%	86.32%	93.85%	98.98%	81.11%	75.00%	90.00%	85.07%	93.44%	98.97%	81.82%	50.00%
16	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m1	jhu-clsp/mmBERT-base	0.155	m1	96.49%	85.93%	86.15%	94.90%	83.33%	83.33%	95.52%	84.86%	86.07%	94.85%	84.09%	66.67%
17	e2_qwen3_17b_base_genlabel_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.081	m1	94.43%	87.09%	91.54%	98.98%	86.67%	91.67%	92.93%	86.14%	90.98%	98.97%	87.50%	83.33%
18	e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-1.7B-Base	0.105	m1	94.48% ±0.80	86.77% ±0.78	93.33% ±0.44	98.30% ±1.56	89.26% ±1.70	91.67% ±0.00	93.07% ±1.00	85.79% ±1.25	92.90% ±0.47	98.28% ±1.57	89.77% ±1.97	83.33% ±0.00
19	e3_ensemble_top_greedy_m1	–	0.113	m1	95.29%	86.90%	94.62%	98.98%	88.89%	91.67%	94.05%	85.93%	95.08%	98.97%	89.77%	83.33%
20	e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m1	intfloat/multilingual-e5-large-instruct	0.017	m1	94.85% ±0.59	87.48% ±2.65	94.36% ±0.44	97.62% ±0.59	84.44% ±1.92	83.33% ±0.00	93.53% ±0.75	86.57% ±2.72	94.54% ±0.95	97.59% ±0.60	85.23% ±1.97	66.67% ±0.00
21	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1	jhu-clsp/mmBERT-base	0.108	m1	92.60% ±1.37	84.65% ±3.28	92.56% ±5.56	96.26% ±0.59	82.59% ±5.25	83.33% ±14.43	91.49% ±1.39	83.51% ±3.63	92.08% ±5.93	96.22% ±0.60	83.33% ±5.37	66.67% ±28.87
22	e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1	Qwen/Qwen3-4B-Base	0.056	m1	95.54% ±0.48	84.59% ±0.67	95.38% ±0.77	98.64% ±0.59	91.11% ±1.11	80.56% ±4.81	94.54% ±0.55	83.44% ±0.44	95.08% ±0.82	98.63% ±0.60	91.29% ±1.31	61.11% ±9.62
23	e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1	unsloth/gemma-3-1b-pt	0.045	m1	94.43% ±0.28	85.45% ±0.14	95.00% ±0.54	95.92% ±1.44	86.67% ±6.29	75.00% ±0.00	93.02% ±0.24	84.33% ±0.15	95.49% ±0.58	95.88% ±1.46	87.50% ±6.43	50.00% ±0.00
24	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepeval_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1, nhãn m1	vinai/phobert-base-v2	0.033	m1	94.01% ±0.57	84.07% ±1.06	91.79% ±1.60	98.64% ±0.59	86.30% ±2.57	66.67% ±8.33	92.76% ±0.87	83.08% ±1.25	91.80% ±1.42	98.63% ±0.60	87.12% ±2.62	33.33% ±16.67
25	e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1	unsloth/gemma-3-1b-pt	0.045	m1	93.93% ±1.11	83.43% ±1.53	94.10% ±0.89	95.58% ±1.18	88.52% ±3.57	80.56% ±4.81	92.39% ±1.50	82.16% ±1.61	94.54% ±0.95	95.53% ±1.19	89.39% ±3.65	61.11% ±9.62
26	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1	jhu-clsp/mmBERT-base	0.052	m1	94.21% ±0.47	85.36% ±2.84	91.03% ±1.60	95.24% ±2.95	83.70% ±5.48	91.67% ±8.33	93.19% ±0.38	84.15% ±3.05	90.98% ±1.42	95.19% ±2.98	84.47% ±5.61	83.33% ±16.67
27	e2_mmbert_base_cross_ce_synth1x_frozenemb_lr2e5_fbmatched_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	jhu-clsp/mmBERT-base	0.051	m1	94.96% ±0.73	86.64% ±2.35	91.79% ±2.91	95.92% ±1.02	84.81% ±2.80	86.11% ±4.81	93.76% ±0.69	85.64% ±2.61	91.80% ±2.84	95.88% ±1.03	85.61% ±2.86	72.22% ±9.62
28	e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.118	m1	95.05% ±1.03	84.26% ±2.25	92.82% ±3.79	95.92% ±1.02	85.56% ±2.94	83.33% ±8.33	93.76% ±1.37	83.16% ±2.37	92.90% ±3.87	95.88% ±1.03	86.36% ±3.01	66.67% ±16.67
29	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1	vinai/phobert-base-v2	0.008	m1	95.03%	85.74%	93.08%	96.94%	85.56%	16.67%	93.88%	84.86%	94.26%	96.91%	86.36%	16.67%
30	e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1	Qwen/Qwen3-4B-Base	0.162	m1	95.29%	81.12%	96.15%	100.00%	92.22%	100.00%	94.22%	79.53%	95.90%	100.00%	93.18%	100.00%
31	e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1	Qwen/Qwen3-8B-Base	0.170	m1	96.36%	87.09%	92.31%	98.98%	92.22%	91.67%	95.43%	86.35%	91.80%	98.97%	93.18%	83.33%
32	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_m1	vinai/phobert-base-v2	0.073	m1	95.43% ±0.34	82.02% ±0.11	94.87% ±0.44	98.64% ±1.56	89.63% ±1.28	80.56% ±12.73	94.40% ±0.09	80.81% ±0.37	95.08% ±0.00	98.63% ±1.57	90.53% ±1.31	61.11% ±25.46
33	e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	95.07% ±0.59	85.68% ±0.59	93.33% ±2.91	96.94% ±1.02	86.30% ±6.79	75.00% ±0.00	93.82% ±0.72	84.72% ±0.69	93.44% ±2.84	96.91% ±1.03	87.12% ±6.94	50.00% ±0.00
34	e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-base	0.031	m1	93.86% ±0.89	86.19% ±2.32	95.13% ±1.78	98.30% ±0.59	84.44% ±2.94	63.89% ±4.81	92.24% ±1.12	85.00% ±2.56	94.81% ±1.89	98.28% ±0.60	85.23% ±3.01	27.78% ±9.62
35	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.061	m1	94.19% ±0.63	84.84% ±3.23	93.59% ±1.18	96.60% ±1.18	83.33% ±5.09	77.78% ±4.81	92.90% ±0.55	83.72% ±3.45	93.72% ±1.25	96.56% ±1.19	84.09% ±5.21	55.56% ±9.62
36	e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1	Qwen/Qwen3-4B-Base	0.131	m1	96.20% ±1.10	85.81% ±1.28	94.36% ±2.70	98.98% ±1.02	89.63% ±1.70	86.11% ±4.81	95.29% ±1.34	84.86% ±1.49	93.99% ±2.88	98.97% ±1.03	90.53% ±1.74	72.22% ±9.62
37	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.043	m1	94.50% ±0.50	84.59% ±0.51	92.05% ±2.47	97.96% ±1.02	87.04% ±4.49	72.22% ±9.62	93.05% ±0.56	83.87% ±0.33	91.80% ±2.17	97.94% ±1.03	87.88% ±4.59	44.44% ±19.25
38	e2_gemma3_1b_pt_cross_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.053	m1	94.83% ±0.53	85.36% ±1.50	93.85% ±0.00	95.92% ±1.02	84.81% ±2.57	75.00% ±0.00	93.59% ±0.61	84.58% ±1.57	94.54% ±0.47	95.88% ±1.03	85.61% ±2.62	50.00% ±0.00
39	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1	vinai/phobert-base-v2	0.037	m1	94.92% ±1.12	84.20% ±1.84	92.82% ±1.78	97.28% ±0.59	87.78% ±1.92	83.33% ±8.33	93.74% ±1.39	83.08% ±1.91	92.90% ±2.37	97.25% ±0.60	88.64% ±1.97	66.67% ±16.67
40	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m1	vinai/phobert-base-v2	0.085	m1	95.85% ±0.45	84.78% ±2.01	91.54% ±0.77	98.98% ±0.00	87.78% ±1.92	80.56% ±12.73	94.91% ±0.52	83.65% ±2.18	91.80% ±0.82	98.97% ±0.00	88.64% ±1.97	61.11% ±25.46
41	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_m1	vinai/phobert-base-v2	0.028	m1	94.74% ±0.97	83.24% ±2.03	93.08% ±1.33	97.62% ±0.59	85.56% ±1.11	80.56% ±20.97	93.39% ±1.26	82.30% ±2.25	93.17% ±0.95	97.59% ±0.60	86.36% ±1.14	61.11% ±41.94
42	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1	jhu-clsp/mmBERT-base	0.082	m1	94.88% ±0.62	84.84% ±3.17	91.79% ±3.87	95.92% ±1.02	82.59% ±3.39	80.56% ±4.81	93.51% ±0.86	83.72% ±3.23	91.80% ±3.76	95.88% ±1.03	83.33% ±3.47	61.11% ±9.62
43	e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.057	m1	94.04% ±0.40	84.59% ±1.67	92.56% ±2.35	95.58% ±2.57	83.70% ±3.39	88.89% ±4.81	92.67% ±0.40	83.51% ±2.09	92.62% ±2.17	95.53% ±2.59	84.47% ±3.47	77.78% ±9.62
44	e2_qwen3_17b_base_cross_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.101	m1	94.35% ±0.77	85.04% ±0.68	91.54% ±2.77	97.96% ±2.04	88.89% ±2.22	80.56% ±9.62	92.99% ±0.92	83.87% ±0.75	91.80% ±2.96	97.94% ±2.06	89.77% ±2.27	61.11% ±19.25
45	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m1	vinai/phobert-base-v2	0.052	m1	94.61% ±0.83	84.78% ±3.93	93.33% ±1.18	97.96% ±1.02	87.78% ±2.22	75.00% ±14.43	93.28% ±0.96	83.80% ±4.15	93.17% ±1.71	97.94% ±1.03	88.26% ±1.74	50.00% ±28.87
46	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1	vinai/phobert-base-v2	0.077	m1	95.34% ±0.84	84.52% ±3.08	94.10% ±0.44	98.64% ±0.59	90.00% ±2.22	72.22% ±9.62	94.17% ±1.13	83.65% ±3.21	94.26% ±0.82	98.63% ±0.60	90.91% ±2.27	44.44% ±19.25
47	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	vinai/phobert-base-v2	0.058	m1	94.61% ±0.15	84.65% ±0.78	92.31% ±0.77	97.28% ±2.12	87.41% ±3.39	80.56% ±12.73	93.22% ±0.18	83.72% ±0.96	92.62% ±0.00	97.25% ±2.15	88.26% ±3.47	61.11% ±25.46
48	e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1	unsloth/gemma-3-1b-pt	0.056	m1	94.92% ±0.44	83.62% ±1.86	93.08% ±0.77	96.60% ±1.18	84.81% ±3.90	80.56% ±4.81	93.68% ±0.65	82.52% ±1.90	93.72% ±1.71	96.56% ±1.19	85.61% ±3.99	61.11% ±9.62
49	e2_gemma3_270m_cross_ce_real_lora_m1gemma-3-270m · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn m1	unsloth/gemma-3-270m	0.098	m1	94.19% ±0.57	81.89% ±1.73	90.77% ±1.33	95.58% ±0.59	80.37% ±6.12	83.33% ±8.33	92.84% ±0.99	80.45% ±1.92	91.26% ±1.25	95.53% ±0.60	81.06% ±6.26	66.67% ±16.67
50	e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-large-instruct	0.029	m1	95.10% ±0.57	86.90% ±1.02	93.08% ±1.54	98.64% ±0.59	84.07% ±2.31	72.22% ±9.62	93.97% ±0.69	86.00% ±1.17	93.72% ±1.25	98.63% ±0.60	84.85% ±2.37	44.44% ±19.25
51	e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.034	m1	92.87% ±1.66	86.00% ±1.57	94.10% ±1.18	97.28% ±1.18	84.07% ±1.70	94.44% ±4.81	92.24% ±0.95	85.00% ±1.61	93.99% ±2.06	97.25% ±1.19	84.85% ±1.74	88.89% ±9.62
52	e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.051	m1	93.55% ±0.23	86.58% ±0.95	93.08% ±0.77	93.88% ±5.40	82.59% ±2.57	88.89% ±12.73	92.47% ±0.22	85.57% ±1.05	93.72% ±0.95	93.81% ±5.46	83.33% ±2.62	77.78% ±25.46
53	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1	jhu-clsp/mmBERT-base	0.040	m1	93.81% ±0.94	84.52% ±2.19	94.10% ±0.89	95.92% ±2.70	81.85% ±5.59	72.22% ±4.81	92.24% ±1.12	83.37% ±2.16	93.99% ±0.47	95.88% ±2.73	82.58% ±5.72	44.44% ±9.62
54	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1	jhu-clsp/mmBERT-base	0.057	m1	94.30% ±1.48	85.29% ±0.29	93.59% ±1.18	96.60% ±1.56	84.07% ±2.31	83.33% ±0.00	93.13% ±1.22	84.08% ±0.44	93.72% ±1.25	96.56% ±1.57	84.85% ±2.37	66.67% ±0.00
55	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.070	m1	95.10%	86.90%	92.31%	96.94%	83.33%	91.67%	94.05%	85.93%	92.62%	96.91%	84.09%	83.33%
56	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.047	m1	94.39% ±0.77	85.61% ±0.87	91.28% ±0.89	96.26% ±2.57	84.07% ±3.21	75.00% ±8.33	93.02% ±1.02	84.58% ±0.86	91.53% ±0.95	96.22% ±2.59	84.85% ±3.28	50.00% ±16.67
57	e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1	Qwen/Qwen3-1.7B-Base	0.095	m1	95.45% ±0.27	86.77% ±3.00	91.28% ±1.94	96.94% ±1.77	88.52% ±2.31	80.56% ±12.73	94.22% ±0.31	85.57% ±3.20	90.98% ±1.64	96.91% ±1.79	89.39% ±2.37	66.67% ±16.67
58	e2_xlmr_base_khonly_ce_real_frozenemb_m1XLM-R base · chỉ câu KH · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1	FacebookAI/xlm-roberta-base	0.012	m1	93.64%	83.04%	94.62%	98.98%	87.78%	66.67%	91.72%	82.09%	95.08%	98.97%	88.64%	33.33%
59	e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.042	m1	94.19% ±0.89	84.71% ±1.16	93.85% ±2.04	95.92% ±1.77	82.22% ±1.92	80.56% ±9.62	92.99% ±0.95	83.65% ±1.29	94.26% ±2.17	95.88% ±1.79	82.95% ±1.97	61.11% ±19.25
60	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m1	vinai/phobert-base-v2	0.024	m1	94.37% ±0.13	83.24% ±1.45	94.10% ±1.18	99.66% ±0.59	86.30% ±6.51	72.22% ±9.62	92.87% ±0.25	82.37% ±1.50	93.99% ±0.95	99.66% ±0.60	87.12% ±6.66	44.44% ±19.25
61	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m1	vinai/phobert-base-v2	0.031	m1	95.34% ±0.60	82.79% ±2.66	92.05% ±2.22	97.96% ±1.77	85.19% ±6.42	75.00% ±14.43	94.17% ±0.78	81.73% ±2.75	92.35% ±2.37	97.94% ±1.79	85.98% ±6.56	50.00% ±28.87
62	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m1	vinai/phobert-base-v2	0.031	m1	94.94% ±0.96	84.71% ±0.59	94.10% ±0.44	97.62% ±1.18	84.44% ±3.33	75.00% ±8.33	93.62% ±1.29	83.87% ±0.54	94.26% ±0.82	97.59% ±1.19	85.23% ±3.41	50.00% ±16.67
63	e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1	unsloth/gemma-3-1b-pt	0.096	m1	94.79% ±1.00	85.04% ±2.12	90.77% ±0.00	96.60% ±1.18	87.41% ±3.39	55.56% ±26.79	93.51% ±1.33	84.15% ±2.57	90.71% ±0.47	96.56% ±1.19	87.88% ±3.65	38.89% ±9.62
64	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1	jhu-clsp/mmBERT-base	0.061	m1	92.98% ±0.50	85.29% ±1.98	93.85% ±2.04	96.26% ±3.12	84.44% ±3.33	80.56% ±20.97	91.78% ±0.43	84.29% ±2.15	94.26% ±2.17	96.22% ±3.15	85.23% ±3.41	61.11% ±41.94
65	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.017	m1	95.12% ±0.19	85.48% ±1.18	92.56% ±1.78	97.62% ±0.59	86.30% ±2.31	58.33% ±0.00	93.88% ±0.31	84.51% ±1.42	93.17% ±1.71	97.59% ±0.60	87.12% ±2.37	16.67% ±0.00
66	e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1	vinai/phobert-base-v2	0.026	m1	95.56% ±0.60	86.19% ±1.62	90.77% ±1.33	98.30% ±0.59	85.93% ±0.64	66.67% ±8.33	94.48% ±0.65	85.29% ±1.67	91.80% ±1.42	98.28% ±0.60	86.74% ±0.66	33.33% ±16.67
67	e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.065	m1	94.88% ±0.66	84.20% ±1.54	92.56% ±1.18	97.28% ±1.56	86.30% ±0.64	77.78% ±4.81	93.62% ±0.91	83.23% ±1.60	92.62% ±0.82	97.25% ±1.57	86.74% ±1.31	55.56% ±9.62
68	e2_xlmr_base_cross_logitadjust_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.018	m1	93.66% ±0.62	81.18% ±1.57	93.59% ±1.18	98.98% ±0.00	87.41% ±2.57	75.00% ±0.00	91.87% ±0.75	79.96% ±1.49	93.99% ±1.89	98.97% ±0.00	87.88% ±2.37	50.00% ±0.00
69	e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-large-instruct	0.026	m1	94.94% ±1.00	89.08% ±0.73	92.05% ±1.60	97.28% ±1.56	85.93% ±1.28	83.33% ±0.00	93.62% ±1.35	88.27% ±0.74	91.80% ±2.17	97.25% ±1.57	86.74% ±1.31	66.67% ±0.00
70	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1	jhu-clsp/mmBERT-base	0.072	m1	95.23% ±0.91	86.26% ±1.93	89.23% ±2.66	95.92% ±2.70	81.48% ±6.70	75.00% ±8.33	93.97% ±1.10	85.29% ±2.06	89.34% ±2.84	95.88% ±2.73	82.20% ±6.85	50.00% ±16.67
71	e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1	jhu-clsp/mmBERT-base	0.082	m1	94.81% ±0.68	86.06% ±1.75	93.08% ±1.54	94.56% ±2.57	81.11% ±7.78	77.78% ±9.62	93.51% ±0.82	84.93% ±1.84	93.44% ±1.64	94.50% ±2.59	81.82% ±7.95	55.56% ±19.25
72	e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.067	m1	93.17%	82.85%	93.08%	97.96%	85.56%	83.33%	91.55%	81.66%	93.44%	97.94%	86.36%	66.67%
73	e2_visobert_cross_ce_real_frozenemb_lr4e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	uitnlp/visobert	0.037	m1	90.85%	82.08%	93.08%	93.88%	84.44%	75.00%	89.22%	80.60%	93.44%	93.81%	85.23%	50.00%
74	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1	FacebookAI/xlm-roberta-base	0.027	m1	92.45%	86.51%	89.23%	98.98%	85.56%	75.00%	90.78%	85.29%	90.16%	98.97%	86.36%	50.00%
75	e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m1mmBERT base · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.057	m1	94.41% ±0.54	85.68% ±3.07	93.59% ±1.94	96.26% ±0.59	83.33% ±3.33	66.67% ±8.33	93.28% ±0.38	84.79% ±3.35	93.72% ±1.71	96.22% ±0.60	84.09% ±3.41	33.33% ±16.67
76	e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.206	m1	96.09% ±1.26	86.58% ±3.38	88.72% ±5.00	98.30% ±2.12	88.15% ±0.64	86.11% ±9.62	95.11% ±1.59	85.57% ±3.63	88.80% ±4.94	98.28% ±2.15	89.02% ±0.66	72.22% ±19.25
77	e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.034	m1	95.47% ±0.27	85.36% ±0.69	90.51% ±1.18	99.32% ±0.59	88.52% ±1.28	75.00% ±14.43	94.22% ±0.38	84.43% ±0.98	90.71% ±1.25	99.31% ±0.60	88.64% ±1.97	50.00% ±28.87
78	e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.052	m1	95.91% ±1.12	84.26% ±2.11	91.54% ±4.28	98.64% ±0.59	83.33% ±5.09	66.67% ±0.00	94.80% ±1.40	83.23% ±2.16	91.53% ±4.66	98.63% ±0.60	84.09% ±5.21	33.33% ±0.00
79	e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.033	m1	95.69% ±0.74	84.14% ±2.39	92.56% ±2.47	97.96% ±0.00	82.59% ±5.25	66.67% ±8.33	94.68% ±1.01	83.23% ±2.54	92.62% ±2.84	97.94% ±0.00	83.33% ±5.37	33.33% ±16.67
80	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1	jhu-clsp/mmBERT-base	0.144	m1	94.19% ±0.55	86.32% ±0.84	91.54% ±0.77	93.54% ±2.36	85.56% ±1.92	75.00% ±8.33	92.61% ±0.64	85.36% ±0.89	91.53% ±1.25	93.47% ±2.38	86.36% ±1.97	50.00% ±16.67
81	e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.082	m1	94.23% ±1.50	85.10% ±1.95	91.03% ±4.64	98.98% ±1.02	83.70% ±2.80	97.22% ±4.81	92.90% ±1.86	84.01% ±2.03	90.98% ±4.56	98.97% ±1.03	84.47% ±2.86	94.44% ±9.62
82	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1	jhu-clsp/mmBERT-base	0.108	m1	95.54% ±0.88	86.64% ±1.13	91.79% ±2.35	96.26% ±1.56	82.59% ±3.57	80.56% ±4.81	94.37% ±1.07	85.71% ±1.13	92.08% ±2.37	96.22% ±1.57	83.71% ±4.30	61.11% ±9.62
83	e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-0.6B-Base	0.075	m1	94.12% ±1.24	83.94% ±2.17	93.59% ±1.60	98.30% ±0.59	89.63% ±2.31	86.11% ±4.81	92.73% ±1.66	82.59% ±2.28	93.44% ±1.64	98.28% ±0.60	90.53% ±2.37	72.22% ±9.62
84	e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1	Qwen/Qwen3-0.6B-Base	0.077	m1	92.51% ±0.93	83.30% ±3.61	90.26% ±3.87	96.26% ±1.18	87.41% ±2.80	77.78% ±9.62	91.61% ±2.03	82.02% ±3.87	90.44% ±4.13	96.22% ±1.19	88.26% ±2.86	55.56% ±19.25
85	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_m1	vinai/phobert-base-v2	0.093	m1	94.54% ±1.60	87.15% ±3.58	92.56% ±0.44	99.66% ±0.59	85.93% ±1.28	88.89% ±12.73	93.16% ±1.93	86.28% ±3.73	92.62% ±0.82	99.66% ±0.60	86.74% ±1.31	77.78% ±25.46
86	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m1	vinai/phobert-base-v2	0.104	m1	95.52% ±1.14	85.93% ±2.04	91.54% ±3.08	98.98% ±1.02	83.33% ±3.33	77.78% ±9.62	94.37% ±1.43	85.07% ±2.06	90.98% ±3.28	98.97% ±1.03	84.09% ±3.41	55.56% ±19.25
87	e2_xlmr_base_cross_ce_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.044	m1	94.01% ±1.01	83.30% ±4.07	90.26% ±3.55	97.28% ±1.18	87.78% ±1.11	69.44% ±12.73	92.47% ±1.26	82.16% ±4.72	90.98% ±3.57	97.25% ±1.19	89.02% ±0.66	38.89% ±25.46
88	e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.019	m1	94.08% ±0.96	82.85% ±1.20	92.56% ±3.20	98.64% ±0.59	86.30% ±2.80	83.33% ±8.33	92.41% ±1.21	81.52% ±1.37	93.44% ±2.46	98.63% ±0.60	87.12% ±2.86	66.67% ±16.67
89	e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.049	m1	94.72% ±0.28	83.62% ±1.84	93.08% ±1.33	97.62% ±1.56	86.30% ±1.70	77.78% ±9.62	93.42% ±0.40	82.66% ±2.05	93.17% ±1.25	97.94% ±1.03	87.12% ±1.74	55.56% ±19.25
90	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.159	m1	95.32% ±0.68	84.20% ±2.78	91.03% ±1.60	98.98% ±1.02	84.44% ±5.09	61.11% ±31.55	94.20% ±1.04	83.16% ±2.93	91.53% ±1.89	98.97% ±1.03	85.23% ±5.21	50.00% ±16.67
91	e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-4B	0.061	m1	93.94% ±0.52	81.02% ±2.32	90.00% ±3.26	88.78% ±8.66	78.89% ±4.71	83.33% ±11.79	92.24% ±0.61	79.64% ±2.26	89.34% ±3.48	88.66% ±8.75	78.98% ±4.02	66.67% ±23.57
92	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.025	m1	95.83%	84.59%	89.23%	97.96%	82.22%	16.67%	94.66%	83.80%	90.16%	97.94%	82.95%	16.67%
93	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.152	m1	94.23% ±0.53	83.82% ±3.36	84.10% ±0.44	93.88% ±0.00	79.63% ±4.21	86.11% ±4.81	92.67% ±0.69	82.44% ±3.65	83.33% ±0.47	93.81% ±0.00	80.30% ±4.30	72.22% ±9.62
94	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m1	vinai/phobert-base-v2	0.035	m1	94.61% ±1.06	84.84% ±2.31	93.59% ±1.60	98.30% ±1.56	87.41% ±2.80	77.78% ±4.81	93.22% ±1.26	83.80% ±2.67	93.72% ±1.89	98.28% ±1.57	88.26% ±2.86	55.56% ±9.62
95	e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-base	0.052	m1	93.57%	84.97%	90.00%	96.94%	83.33%	25.00%	91.90%	83.80%	90.16%	96.91%	84.09%	33.33%
96	e2_visobert_cross_ce_real_frozenemb_lr2e5_m1	uitnlp/visobert	0.027	m1	92.18%	79.96%	90.00%	94.90%	86.67%	58.33%	90.60%	78.25%	90.98%	94.85%	87.50%	16.67%
97	e2_visobert_cross_ce_real_frozenemb_lr1e5_m1	uitnlp/visobert	0.186	m1	92.31%	82.08%	86.15%	94.90%	76.67%	66.67%	90.52%	81.02%	86.07%	94.85%	76.14%	33.33%
98	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1	FacebookAI/xlm-roberta-base	0.161	m1	93.77%	83.82%	90.00%	97.96%	82.22%	58.33%	91.98%	82.94%	90.98%	97.94%	82.95%	16.67%
99	e1_tfidf_hybrid_qwen3emb_lr_m1TF-IDF câu KH (char 1-5, word 1-2-gram, giữ từ 1 ký tự, thêm kênh char bỏ dấu) + câu bot (char 2-4) + tương tác loại câu bot × từ câu KH · vector đóng băng Qwen3-Embedding-0.6B (cặp bot + KH, pooling last_token), trọng số dense 1 · LogReg · train thật, nhãn m1	Qwen/Qwen3-Embedding-0.6B	0.149	m1	94.37%	79.58%	84.62%	89.80%	78.89%	66.67%	93.19%	78.04%	84.43%	89.69%	79.55%	33.33%
100	e1_tfidf_svmcal_ce_synth1x_fbmatched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	TF-IDF	0.207	m1	95.10%	84.78%	81.54%	94.90%	78.89%	66.67%	93.79%	83.80%	81.15%	94.85%	79.55%	33.33%
101	e1_tfidf_svmcal_ce_synth1x_nofb_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent omi_intent_repeat, callback, relationship, negate, affirm), nhãn m1	TF-IDF	0.164	m1	95.03%	84.39%	80.77%	94.90%	78.89%	66.67%	93.71%	83.37%	80.33%	94.85%	79.55%	33.33%
102	e1_tfidf_svmcal_ce_synth1x_noqparticle_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm q_particle), nhãn m1	TF-IDF	0.280	m1	93.24%	82.66%	84.62%	94.90%	77.78%	83.33%	91.72%	81.45%	84.43%	94.85%	78.41%	66.67%
103	e1_tfidf_svmcal_ce_real_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn m1	TF-IDF	0.178	m1	94.43%	85.36%	83.85%	94.90%	78.89%	66.67%	93.02%	84.22%	83.61%	94.85%	79.55%	33.33%
104	e1_tfidf_svmcal_ce_synth1x_nofiller_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm filler_only), nhãn m1	TF-IDF	0.267	m1	93.51%	81.70%	87.69%	94.90%	76.67%	75.00%	91.64%	80.38%	87.70%	94.85%	77.27%	50.00%
105	e1_tfidf_svmcal_ce_synth1x_noofftopic_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm prefix_offtopic), nhãn m1	TF-IDF	0.298	m1	93.37%	82.85%	84.62%	94.90%	78.89%	83.33%	92.07%	81.66%	84.43%	94.85%	79.55%	66.67%
106	e2_dual_qwen3emb06b_concat_ce_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.227	m1	93.11% ±1.40	77.65% ±0.58	85.90% ±1.18	91.84% ±3.06	68.52% ±6.79	66.67% ±8.33	91.44% ±1.69	76.05% ±0.54	85.52% ±0.95	91.75% ±3.09	68.94% ±6.94	33.33% ±16.67
107	e1_tfidf_svmcal_ce_synth1x_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1, nhãn m1	TF-IDF	0.282	m1	93.31%	82.08%	83.85%	94.90%	77.78%	83.33%	91.72%	80.81%	83.61%	94.85%	78.41%	66.67%
108	e1_tfidf_svmcal_ce_synth1x_matched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ intent của train thật), nhãn m1	TF-IDF	0.190	m1	94.50%	85.36%	83.08%	94.90%	77.78%	66.67%	93.02%	84.22%	82.79%	94.85%	78.41%	33.33%
109	e1_tfidf_svmcal_ce_synth1x_nomoney_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm money), nhãn m1	TF-IDF	0.310	m1	93.31%	83.04%	85.38%	94.90%	80.00%	83.33%	91.64%	81.88%	85.25%	94.85%	80.68%	66.67%
110	e1_tfidf_svmcal_ce_synth1x_fbonly_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent nlu_fallback), nhãn m1	TF-IDF	0.356	m1	91.98%	82.85%	85.38%	94.90%	78.89%	83.33%	90.60%	81.45%	85.25%	94.85%	79.55%	66.67%
–	e2_bamibert_cross_ce_real_frozenemb_lr1e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_bamibert_cross_ce_real_frozenemb_lr2e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_bamibert_cross_ce_real_frozenemb_lr4e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e3_fbverifier_devtop1_familygeomean_fbf1_m1	–	hard	m1	97.81%	92.10%	85.38%	100.00%	88.89%	91.67%	97.33%	91.68%	86.07%	100.00%	89.77%	83.33%
–	e3_fbverifier_devtop1_familygeomean_macrof1_m1	–	hard	m1	97.81%	92.10%	85.38%	100.00%	88.89%	91.67%	97.33%	91.68%	86.07%	100.00%	89.77%	83.33%
1	e3_ensemble_mean_all	–	0.065	v1	93.91%	83.16%	88.78%	90.54%	82.19%	60.00%	91.83%	81.56%	90.80%	91.67%	83.10%	50.00%
2	e3_ensemble_greedy_all	–	0.070	v1	93.81%	83.42%	87.76%	91.89%	80.82%	60.00%	91.83%	81.84%	89.66%	93.06%	81.69%	50.00%
3	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5	–	0.023	v1	94.72%	84.72%	82.65%	90.54%	78.08%	60.00%	92.94%	83.57%	83.91%	91.67%	78.87%	50.00%
4	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1	vinai/phobert-base-v2	0.011	v1	95.33%	79.53%	82.65%	91.89%	83.56%	40.00%	93.91%	77.81%	86.21%	93.06%	84.51%	25.00%
5	e2_xlmr_base_cross_wcesqrt_real_frozenemb	FacebookAI/xlm-roberta-base	0.021	v1	93.81% ±0.10	84.54% ±1.43	84.35% ±1.56	92.34% ±0.78	83.11% ±2.09	60.00% ±20.00	91.74% ±0.21	83.29% ±1.26	86.21% ±1.99	93.52% ±0.80	84.04% ±2.15	50.00% ±25.00
6	e3_prior_em_mmbert_bottype	–	0.082	v1	95.23%	60.36%	82.65%	85.14%	75.34%	40.00%	93.35%	55.62%	83.91%	86.11%	76.06%	25.00%
7	e3_prior_em_mmbert_global	–	0.082	v1	95.84%	79.79%	85.71%	89.19%	75.34%	40.00%	93.91%	78.67%	89.66%	90.28%	76.06%	25.00%
8	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5	jhu-clsp/mmBERT-base	0.061	v1	94.48% ±0.16	82.99% ±2.35	85.71% ±1.02	89.19% ±1.35	83.56% ±4.94	86.67% ±11.55	92.71% ±0.29	81.65% ±2.61	88.51% ±1.15	90.28% ±1.39	84.51% ±5.08	83.33% ±14.43
9	e2_xlmr_base_cross_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	94.45% ±0.12	85.66% ±1.50	82.99% ±3.58	91.44% ±0.78	81.28% ±2.85	60.00% ±0.00	92.66% ±0.24	84.63% ±1.66	84.67% ±4.04	92.59% ±0.80	82.16% ±2.93	50.00% ±0.00
10	e2_xlmr_base_cross_logitadjust_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	93.74% ±0.61	83.85% ±0.91	85.37% ±4.12	91.89% ±1.35	83.11% ±2.09	66.67% ±11.55	91.60% ±0.83	82.61% ±1.01	87.36% ±3.04	93.06% ±1.39	84.04% ±2.15	58.33% ±14.43
11	e2_qwen3_06b_base_cross_ce_real_lora	Qwen/Qwen3-0.6B-Base	0.098	v1	93.30%	84.46%	84.69%	90.54%	82.19%	60.00%	91.27%	83.00%	87.36%	91.67%	83.10%	50.00%
12	e3_ensemble_stack_all	–	0.088	v1	92.99%	87.05%	84.69%	87.84%	79.45%	60.00%	90.58%	85.88%	87.36%	88.89%	80.28%	75.00%
13	e2_qwen3_17b_base_cross_ce_real_lora	Qwen/Qwen3-1.7B-Base	0.083	v1	94.01% ±0.27	83.51% ±1.28	83.67% ±2.04	90.09% ±0.78	83.11% ±2.85	86.67% ±23.09	92.20% ±0.21	82.13% ±1.50	85.82% ±1.76	91.20% ±0.80	84.04% ±2.93	83.33% ±28.87
14	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5	vinai/phobert-base-v2	0.022	v1	94.72% ±0.27	84.72% ±1.13	84.35% ±1.56	91.44% ±0.78	82.19% ±2.37	46.67% ±11.55	93.03% ±0.29	83.38% ±1.16	86.59% ±1.33	92.59% ±0.80	83.10% ±2.44	33.33% ±14.43
15	e2_me5_base_cross_ce_real_frozenemb_lr1e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	intfloat/multilingual-e5-base	0.088	v1	93.40%	84.20%	87.76%	89.19%	80.82%	60.00%	91.27%	82.71%	87.36%	90.28%	81.69%	50.00%
16	e2_me5_base_cross_ce_real_frozenemb_lr4e5	–	0.028	v1	94.52%	84.72%	85.71%	91.89%	80.82%	60.00%	92.80%	83.29%	87.36%	93.06%	81.69%	50.00%
17	e2_xlmr_base_khonly_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.076	v1	94.08% ±0.76	84.46% ±2.07	82.99% ±3.12	91.89% ±1.35	82.65% ±3.16	53.33% ±23.09	92.06% ±1.04	83.19% ±2.31	85.44% ±3.51	93.06% ±1.39	83.57% ±3.25	41.67% ±28.87
18	e2_me5_base_cross_ce_real_frozenemb_lr2e5	intfloat/multilingual-e5-base	0.056	v1	93.67% ±0.52	86.79% ±0.26	84.35% ±2.12	91.89% ±1.35	81.74% ±1.58	46.67% ±11.55	91.55% ±0.63	85.88% ±0.50	85.82% ±2.39	93.06% ±1.39	82.63% ±1.63	33.33% ±14.43
19	e2_gemma3_270m_cross_ce_real_lora	unsloth/gemma-3-270m	0.116	v1	93.37% ±1.31	83.59% ±0.30	85.03% ±2.57	89.19% ±0.00	78.08% ±2.37	66.67% ±11.55	91.32% ±1.49	81.94% ±0.44	88.12% ±1.76	90.28% ±0.00	78.87% ±2.44	58.33% ±14.43
20	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen	Qwen/Qwen3-Embedding-0.6B	0.114	v1	93.37% ±0.29	81.52% ±0.91	80.95% ±3.86	88.29% ±2.06	84.02% ±1.58	80.00% ±0.00	91.41% ±0.60	79.73% ±1.01	82.38% ±3.69	89.35% ±2.12	84.98% ±1.63	75.00% ±0.00
21	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5	–	0.075	v1	94.42%	83.94%	84.69%	89.19%	82.19%	60.00%	92.94%	82.71%	88.51%	90.28%	83.10%	50.00%
22	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5	–	0.026	v1	94.62%	82.12%	84.69%	90.54%	80.82%	40.00%	92.80%	81.27%	86.21%	91.67%	81.69%	50.00%
23	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	FacebookAI/xlm-roberta-base	0.123	v1	94.21%	86.53%	77.55%	94.59%	76.71%	40.00%	92.38%	85.59%	79.31%	95.83%	77.46%	25.00%
24	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5	jhu-clsp/mmBERT-base	0.120	v1	93.96% ±0.36	85.36% ±1.28	80.61% ±2.89	91.22% ±0.96	84.93% ±5.81	60.00% ±0.00	92.24% ±0.20	84.29% ±1.43	83.91% ±3.25	92.36% ±0.98	85.92% ±5.98	50.00% ±0.00
25	e2_visobert_cross_ce_real_frozenemb_lr2e5	–	0.093	v1	93.40%	85.23%	78.57%	87.84%	79.45%	60.00%	91.41%	84.15%	81.61%	88.89%	80.28%	50.00%
26	e2_visobert_cross_ce_real_frozenemb_lr4e5	–	0.066	v1	92.89%	84.20%	80.61%	89.19%	79.45%	80.00%	90.58%	83.00%	82.76%	90.28%	81.69%	75.00%
27	e1_tfidf_svmcal_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn v1	TF-IDF	0.123	v1	94.11%	82.12%	85.71%	89.19%	76.71%	60.00%	92.38%	80.40%	87.36%	90.28%	77.46%	50.00%
28	e1_tfidf_logreg_balanced_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg cân bằng lớp · train thật, nhãn v1	TF-IDF	0.256	v1	93.60%	75.65%	81.63%	87.84%	86.30%	80.00%	91.41%	73.49%	82.76%	88.89%	85.92%	75.00%
29	e1_tfidf_logreg_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật, nhãn v1	TF-IDF	0.125	v1	93.50%	79.27%	84.69%	86.49%	75.34%	60.00%	91.27%	77.23%	86.21%	87.50%	76.06%	50.00%
30	e2_visobert_cross_ce_real_frozenemb_lr1e5	–	0.232	v1	93.81%	83.94%	71.43%	89.19%	78.08%	60.00%	91.97%	83.00%	72.41%	90.28%	78.87%	50.00%
31	e1_tfidf_logreg_ce_synth1x_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật + synth ×1, nhãn v1	TF-IDF	0.234	v1	93.91%	79.53%	79.59%	87.84%	75.34%	60.00%	92.38%	77.52%	81.61%	88.89%	76.06%	50.00%
–	e2_mdeberta_cross_ce_real_frozenemb_lr1e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%
–	e2_videberta_cross_ce_real_frozenemb_lr1e5	–	argmax	v1	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%
–	e2_videberta_cross_ce_real_frozenemb_lr2e5	–	argmax	v1	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%
–	e2_videberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_videberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%
1	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_lv21	jhu-clsp/mmBERT-base	0.076	v2_1	94.18% ±0.85	86.61% ±1.08	78.57% ±2.04	87.84% ±0.00	82.65% ±4.40	66.67% ±30.55	92.57% ±0.76	85.88% ±1.04	81.99% ±1.76	88.89% ±0.00	83.57% ±4.53	58.33% ±38.19
2	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_lv21	vinai/phobert-base-v2	0.033	v2_1	94.82%	87.31%	77.55%	86.49%	76.71%	40.00%	93.07%	86.17%	79.31%	87.50%	77.46%	25.00%

F1 theo từng intent

STT	Model_Name	Backbone	Threshold	Data	clean Affirm_F1	clean Negate_F1	clean Fallback_F1	clean Callback_F1	clean Relation_F1	clean Repeat_F1	reply-unseen Affirm_F1	reply-unseen Negate_F1	reply-unseen Fallback_F1	reply-unseen Callback_F1	reply-unseen Relation_F1	reply-unseen Repeat_F1
0	e0_prod_sft_hard_none_qc_v16_m1 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	m1	97.47%	91.83%	79.50%	98.00%	91.43%	82.76%	97.01%	91.88%	82.46%	97.98%	91.86%	70.59%
0	e0_prod_sft_hard_none_qc_v16 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v1	96.65%	90.98%	76.24%	95.30%	90.00%	76.92%	96.18%	91.48%	81.44%	95.89%	90.51%	72.73%
0	e0_prod_sft_hard_none_qc_v16_lv2 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2	96.56%	90.84%	73.56%	95.17%	92.86%	66.67%	96.04%	91.29%	77.78%	95.10%	94.20%	60.00%
0	e0_prod_sft_hard_none_qc_v16_lv21 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2_1	96.65%	90.98%	76.24%	95.30%	90.00%	76.92%	96.18%	91.48%	81.44%	95.89%	90.51%	72.73%
1	e3_ensemble_devtop1_familygeomean_m1	–	0.051	m1	97.06%	89.99%	68.83%	96.48%	90.29%	78.26%	96.42%	89.21%	68.38%	96.45%	90.70%	54.55%
2	e3_ensemble_devtop1_familygeomean_tempdev_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	97.20%	90.21%	70.17%	97.00%	90.29%	78.26%	96.56%	89.46%	69.57%	96.97%	90.70%	54.55%
3	e3_ensemble_devtop1_familygeomean_temphard_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	97.20%	90.21%	70.17%	97.00%	90.29%	78.26%	96.56%	89.46%	69.57%	96.97%	90.70%	54.55%
4	e3_ensemble_devtop1_familymean_m1	–	0.112	m1	97.14%	90.30%	70.79%	97.00%	91.43%	83.33%	96.48%	89.56%	70.21%	96.97%	91.86%	66.67%
5	e3_ensemble_sweep_k2_m1	–	0.110	m1	97.46%	91.97%	78.96%	94.58%	91.86%	80.00%	96.95%	91.43%	78.50%	94.53%	92.31%	61.54%
6	e3_ensemble_sweep_k3_m1	–	0.051	m1	97.32%	90.96%	73.16%	95.10%	90.70%	72.00%	96.75%	90.30%	72.45%	95.05%	91.12%	46.15%
7	e3_ensemble_sweep_k4_m1	–	0.058	m1	97.48%	90.85%	72.94%	96.04%	91.33%	72.00%	96.93%	90.18%	72.00%	96.00%	91.76%	46.15%
8	e3_ensemble_sweep_k5_m1	–	0.111	m1	97.79%	92.15%	79.11%	96.08%	91.33%	83.33%	97.43%	91.63%	78.93%	96.04%	91.76%	66.67%
9	e3_ensemble_sweep_k6_m1	–	0.087	m1	97.72%	91.71%	76.45%	96.08%	91.33%	86.96%	97.25%	91.13%	76.13%	96.04%	91.76%	72.73%
10	e3_ensemble_sweep_k7_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.087	m1	97.72%	91.71%	76.45%	96.08%	91.33%	86.96%	97.25%	91.13%	76.13%	96.04%	91.76%	72.73%
11	e3_ensemble_sweep_k8_m1	–	0.087	m1	97.72%	91.71%	76.45%	96.08%	91.33%	86.96%	97.25%	91.13%	76.13%	96.04%	91.76%	72.73%
12	e3_ensemble_top_geomean_m1	–	0.073	m1	97.20%	90.71%	72.30%	96.04%	92.57%	76.92%	96.56%	90.02%	71.34%	96.00%	93.02%	57.14%
13	e3_ensemble_top_mean_m1	–	0.073	m1	96.71%	89.54%	65.79%	97.00%	93.10%	76.92%	95.96%	88.84%	65.37%	96.97%	93.57%	57.14%
14	e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	96.05% ±0.69	90.23% ±1.28	66.38% ±6.78	94.87% ±1.19	89.07% ±2.09	70.56% ±4.19	95.14% ±0.81	89.49% ±1.43	65.96% ±6.68	94.81% ±1.20	89.47% ±2.12	36.11% ±12.73
15	e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1	intfloat/multilingual-e5-base	0.017	m1	95.17%	89.51%	62.56%	94.17%	85.88%	64.29%	93.88%	88.57%	61.29%	94.12%	86.23%	37.50%
16	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m1	jhu-clsp/mmBERT-base	0.155	m1	96.87%	89.65%	70.89%	92.54%	87.72%	74.07%	96.14%	88.74%	70.00%	92.46%	88.10%	53.33%
17	e2_qwen3_17b_base_genlabel_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.081	m1	96.54%	90.58%	66.30%	95.10%	89.14%	78.57%	95.65%	89.98%	64.91%	95.05%	89.53%	62.50%
18	e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-1.7B-Base	0.105	m1	96.53% ±0.49	90.73% ±0.23	66.36% ±0.46	95.23% ±1.36	91.11% ±0.44	86.87% ±1.95	95.67% ±0.60	89.97% ±0.48	65.58% ±0.72	95.02% ±1.27	91.32% ±0.68	75.09% ±3.17
19	e3_ensemble_top_greedy_m1	–	0.113	m1	97.06%	90.56%	69.10%	96.52%	92.49%	81.48%	96.38%	89.86%	68.44%	96.48%	92.94%	66.67%
20	e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m1	intfloat/multilingual-e5-large-instruct	0.017	m1	96.49% ±0.23	90.67% ±1.30	70.55% ±5.88	95.03% ±0.50	89.06% ±1.31	75.30% ±5.69	95.66% ±0.31	89.97% ±1.33	69.86% ±5.88	94.98% ±0.50	89.46% ±1.33	55.25% ±7.42
21	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1	jhu-clsp/mmBERT-base	0.108	m1	95.27% ±0.62	88.97% ±1.37	61.31% ±4.48	93.09% ±0.46	86.38% ±2.04	77.59% ±4.48	94.51% ±0.62	88.09% ±1.55	61.77% ±3.97	93.02% ±0.47	86.73% ±2.06	56.86% ±11.89
22	e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1	Qwen/Qwen3-4B-Base	0.056	m1	97.03% ±0.16	89.56% ±0.57	69.10% ±1.38	96.51% ±0.47	92.31% ±0.68	71.58% ±2.43	96.38% ±0.17	88.81% ±0.47	68.33% ±1.37	96.47% ±0.47	92.33% ±0.91	48.73% ±5.35
23	e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1	unsloth/gemma-3-1b-pt	0.045	m1	96.46% ±0.03	89.55% ±0.08	65.70% ±1.12	95.19% ±0.27	89.60% ±2.86	76.91% ±6.94	95.63% ±0.02	88.68% ±0.16	65.19% ±1.56	95.14% ±0.27	90.00% ±2.89	53.08% ±9.79
24	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepeval_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1, nhãn m1	vinai/phobert-base-v2	0.033	m1	96.09% ±0.17	88.59% ±0.73	63.41% ±1.62	95.25% ±1.07	88.42% ±1.93	69.53% ±7.59	95.33% ±0.30	87.79% ±0.84	63.21% ±1.81	95.20% ±1.08	88.97% ±2.09	36.06% ±16.27
25	e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1	unsloth/gemma-3-1b-pt	0.045	m1	96.10% ±0.50	88.15% ±0.45	64.09% ±5.27	95.42% ±0.82	89.34% ±3.03	74.66% ±7.59	95.20% ±0.73	87.24% ±0.54	63.26% ±5.54	95.54% ±0.75	89.90% ±3.20	53.17% ±12.22
26	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1	jhu-clsp/mmBERT-base	0.052	m1	96.09% ±0.21	89.27% ±1.70	65.49% ±2.82	93.01% ±1.02	88.05% ±2.70	71.41% ±4.77	95.40% ±0.21	88.34% ±1.91	65.57% ±2.37	92.94% ±1.03	88.43% ±2.74	53.90% ±3.62
27	e2_mmbert_base_cross_ce_synth1x_frozenemb_lr2e5_fbmatched_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	jhu-clsp/mmBERT-base	0.051	m1	96.55% ±0.32	90.28% ±1.18	67.29% ±2.58	95.11% ±0.76	88.75% ±0.60	77.49% ±3.74	95.80% ±0.30	89.51% ±1.26	66.64% ±2.37	95.06% ±0.76	89.14% ±0.60	59.05% ±6.87
28	e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.118	m1	96.64% ±0.30	88.85% ±1.23	66.47% ±4.32	94.79% ±1.15	88.49% ±1.24	76.05% ±8.16	95.87% ±0.41	87.92% ±1.30	65.77% ±4.64	94.74% ±1.16	88.87% ±1.25	56.07% ±14.15
29	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1	vinai/phobert-base-v2	0.008	m1	96.60%	89.27%	67.60%	94.06%	87.50%	28.57%	95.86%	88.64%	68.86%	94.00%	87.86%	28.57%
30	e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1	Qwen/Qwen3-4B-Base	0.162	m1	96.97%	87.62%	67.02%	96.55%	90.71%	80.00%	96.34%	86.54%	66.10%	96.52%	91.11%	66.67%
31	e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1	Qwen/Qwen3-8B-Base	0.170	m1	97.39%	91.04%	74.07%	96.04%	90.71%	78.57%	96.81%	90.50%	73.20%	96.00%	91.11%	62.50%
32	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_m1	vinai/phobert-base-v2	0.073	m1	96.73% ±0.12	88.10% ±0.15	68.03% ±0.95	95.55% ±0.49	89.03% ±3.00	74.22% ±1.57	96.07% ±0.05	87.16% ±0.35	67.84% ±0.87	95.51% ±0.50	89.40% ±3.06	50.72% ±9.68
33	e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	96.65% ±0.03	89.29% ±0.53	68.75% ±0.38	95.48% ±0.83	89.18% ±3.38	71.15% ±3.33	95.85% ±0.04	88.56% ±0.57	68.20% ±0.35	95.44% ±0.84	89.57% ±3.42	45.24% ±4.12
34	e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-base	0.031	m1	96.12% ±0.44	89.55% ±1.14	67.28% ±3.94	93.83% ±0.54	87.84% ±1.73	66.67% ±5.02	95.14% ±0.57	88.61% ±1.21	66.18% ±4.08	93.77% ±0.55	88.22% ±1.75	30.30% ±10.50
35	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.061	m1	96.05% ±0.28	89.15% ±1.32	65.34% ±4.26	95.64% ±0.95	87.49% ±2.71	72.86% ±5.45	95.23% ±0.31	88.28% ±1.48	64.92% ±4.00	95.59% ±0.96	87.87% ±2.75	49.05% ±8.61
36	e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1	Qwen/Qwen3-4B-Base	0.131	m1	97.35% ±0.52	89.96% ±0.56	73.13% ±5.83	95.91% ±1.47	91.50% ±1.47	75.71% ±9.83	96.76% ±0.65	89.27% ±0.65	72.46% ±5.83	95.87% ±1.48	91.93% ±1.49	57.58% ±13.12
37	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.043	m1	96.26% ±0.20	88.63% ±0.46	65.97% ±1.47	95.37% ±0.51	89.66% ±2.31	73.58% ±6.68	95.35% ±0.25	87.96% ±0.28	65.34% ±1.62	95.32% ±0.52	90.06% ±2.34	45.56% ±10.72
38	e2_gemma3_1b_pt_cross_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.053	m1	96.47% ±0.41	89.28% ±0.98	67.46% ±2.35	95.11% ±0.24	88.41% ±1.29	73.16% ±4.63	95.70% ±0.51	88.60% ±0.95	67.56% ±2.08	95.06% ±0.24	88.79% ±1.31	47.85% ±6.03
39	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1	vinai/phobert-base-v2	0.037	m1	96.48% ±0.55	88.46% ±0.38	67.56% ±2.26	94.24% ±1.35	90.28% ±1.08	82.40% ±2.97	95.71% ±0.72	87.59% ±0.37	67.15% ±2.55	94.19% ±1.37	90.69% ±1.09	64.96% ±2.96
40	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m1	vinai/phobert-base-v2	0.085	m1	96.82% ±0.27	89.24% ±1.24	71.14% ±1.36	94.48% ±0.27	89.12% ±1.52	77.16% ±6.99	96.19% ±0.40	88.35% ±1.32	70.85% ±1.70	94.43% ±0.27	89.67% ±1.29	55.18% ±16.43
41	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_m1	vinai/phobert-base-v2	0.028	m1	96.49% ±0.42	88.07% ±0.57	65.24% ±1.23	95.83% ±0.73	89.02% ±1.00	68.48% ±13.47	95.68% ±0.53	87.32% ±0.70	64.59% ±1.03	95.79% ±0.73	89.41% ±1.02	43.47% ±27.61
42	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1	jhu-clsp/mmBERT-base	0.082	m1	96.44% ±0.21	88.85% ±1.48	67.30% ±4.63	94.32% ±0.22	86.42% ±2.33	68.29% ±1.41	95.55% ±0.33	87.92% ±1.54	66.61% ±4.74	94.26% ±0.23	86.77% ±2.37	44.79% ±2.12
43	e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.057	m1	96.14% ±0.18	89.13% ±1.14	63.28% ±1.50	94.77% ±0.41	87.75% ±0.88	72.92% ±5.08	95.32% ±0.19	88.27% ±1.39	62.90% ±1.75	94.71% ±0.41	88.13% ±0.88	54.17% ±7.22
44	e2_qwen3_17b_base_cross_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.101	m1	96.15% ±0.23	89.31% ±0.41	65.82% ±2.83	95.68% ±2.01	90.73% ±0.99	71.49% ±6.27	95.27% ±0.26	88.46% ±0.46	65.38% ±2.80	95.64% ±2.03	91.14% ±1.01	48.45% ±12.25
45	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m1	vinai/phobert-base-v2	0.052	m1	96.43% ±0.28	88.75% ±1.04	66.97% ±3.14	95.52% ±0.50	89.95% ±0.74	74.78% ±4.67	95.60% ±0.28	87.89% ±1.12	66.35% ±3.35	95.48% ±0.50	90.31% ±0.80	47.77% ±13.38
46	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1	vinai/phobert-base-v2	0.077	m1	96.75% ±0.47	89.02% ±1.77	69.64% ±5.98	94.93% ±0.70	91.01% ±0.64	78.73% ±2.20	96.00% ±0.59	88.31% ±1.84	69.36% ±6.00	94.88% ±0.71	91.42% ±0.65	51.99% ±8.72
47	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	vinai/phobert-base-v2	0.058	m1	96.36% ±0.22	88.34% ±0.09	68.04% ±1.94	95.17% ±0.77	88.88% ±2.40	69.89% ±6.54	95.50% ±0.26	87.58% ±0.23	67.59% ±2.53	95.12% ±0.77	89.26% ±2.44	45.97% ±13.64
48	e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1	unsloth/gemma-3-1b-pt	0.056	m1	96.57% ±0.28	88.51% ±0.72	65.51% ±2.93	94.04% ±0.47	88.90% ±2.52	74.72% ±5.03	95.80% ±0.36	87.75% ±0.71	65.18% ±2.47	93.98% ±0.48	89.30% ±2.56	52.92% ±5.24
49	e2_gemma3_270m_cross_ce_real_lora_m1gemma-3-270m · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn m1	unsloth/gemma-3-270m	0.098	m1	95.78% ±0.33	87.11% ±0.73	63.21% ±1.87	93.20% ±0.29	86.20% ±3.81	66.67% ±0.00	94.81% ±0.55	86.04% ±0.78	62.99% ±2.50	93.45% ±0.04	86.56% ±3.87	44.21% ±7.29
50	e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-large-instruct	0.029	m1	96.65% ±0.34	90.38% ±0.72	68.69% ±0.80	96.03% ±0.98	88.66% ±1.32	71.24% ±1.59	95.98% ±0.41	89.73% ±0.77	68.34% ±0.95	95.99% ±0.98	89.06% ±1.34	43.33% ±5.77
51	e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.034	m1	95.59% ±0.96	89.56% ±0.88	63.43% ±2.22	94.72% ±1.39	88.15% ±0.67	79.39% ±7.59	95.17% ±0.64	88.78% ±0.91	65.03% ±0.49	94.67% ±1.40	88.54% ±0.68	64.77% ±11.30
52	e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.051	m1	95.77% ±0.14	89.93% ±1.03	65.00% ±0.66	94.31% ±2.45	86.76% ±1.97	71.96% ±15.22	95.07% ±0.15	89.19% ±1.16	65.59% ±0.57	94.25% ±2.48	87.12% ±2.00	54.89% ±20.59
53	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1	jhu-clsp/mmBERT-base	0.040	m1	95.98% ±0.68	89.09% ±1.35	63.36% ±3.87	94.79% ±1.09	86.62% ±4.34	66.91% ±4.97	95.00% ±0.81	88.19% ±1.37	62.50% ±3.56	94.73% ±1.10	87.14% ±4.15	38.26% ±6.91
54	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1	jhu-clsp/mmBERT-base	0.057	m1	96.24% ±0.77	89.16% ±0.11	66.15% ±3.66	95.78% ±1.29	88.32% ±1.64	71.17% ±7.64	95.49% ±0.64	88.18% ±0.11	65.98% ±2.71	95.74% ±1.30	88.70% ±1.67	51.43% ±9.90
55	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.070	m1	96.50%	90.47%	70.18%	93.60%	87.72%	75.86%	95.83%	89.66%	70.19%	93.53%	88.10%	58.82%
56	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.047	m1	96.15% ±0.39	89.43% ±0.48	65.36% ±1.66	94.49% ±1.02	88.31% ±1.73	71.48% ±10.61	95.27% ±0.52	88.61% ±0.52	64.90% ±1.69	94.43% ±1.03	88.69% ±1.76	46.12% ±16.90
57	e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1	Qwen/Qwen3-1.7B-Base	0.095	m1	96.91% ±0.12	90.26% ±1.44	67.84% ±2.58	96.61% ±0.80	91.04% ±0.69	81.38% ±4.54	96.10% ±0.18	89.36% ±1.60	66.84% ±2.85	96.57% ±0.81	91.47% ±0.70	66.03% ±5.74
58	e2_xlmr_base_khonly_ce_real_frozenemb_m1XLM-R base · chỉ câu KH · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1	FacebookAI/xlm-roberta-base	0.012	m1	95.63%	88.14%	65.95%	92.38%	89.77%	66.67%	94.49%	87.30%	65.17%	92.31%	90.17%	30.77%
59	e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.042	m1	96.02% ±0.41	89.30% ±0.72	64.84% ±2.12	94.32% ±0.45	86.72% ±1.51	79.29% ±4.89	95.22% ±0.43	88.53% ±0.86	64.89% ±1.75	94.26% ±0.46	87.07% ±1.54	58.66% ±11.29
60	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m1	vinai/phobert-base-v2	0.024	m1	96.28% ±0.16	88.16% ±0.81	64.73% ±0.67	95.91% ±0.52	89.36% ±3.92	76.31% ±3.64	95.37% ±0.15	87.43% ±0.79	64.07% ±0.51	95.87% ±0.53	89.76% ±3.97	48.66% ±11.37
61	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m1	vinai/phobert-base-v2	0.031	m1	96.52% ±0.39	88.27% ±1.48	66.12% ±3.73	95.21% ±0.74	88.39% ±3.32	75.65% ±5.05	95.76% ±0.52	87.50% ±1.54	65.50% ±3.76	95.16% ±0.74	88.77% ±3.37	48.89% ±15.40
62	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m1	vinai/phobert-base-v2	0.031	m1	96.52% ±0.43	88.94% ±0.49	68.09% ±3.61	94.25% ±0.24	88.18% ±2.32	77.00% ±3.80	95.71% ±0.61	88.22% ±0.52	67.52% ±3.87	94.20% ±0.25	88.56% ±2.36	52.03% ±10.99
63	e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1	unsloth/gemma-3-1b-pt	0.096	m1	96.45% ±0.45	88.64% ±1.04	66.83% ±4.82	95.31% ±1.57	89.71% ±2.10	49.97% ±21.22	95.59% ±0.61	87.92% ±1.25	66.82% ±5.27	95.26% ±1.58	89.90% ±2.32	30.00% ±8.66
64	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1	jhu-clsp/mmBERT-base	0.061	m1	95.50% ±0.03	89.36% ±1.03	62.91% ±1.84	94.50% ±1.27	88.18% ±1.84	71.85% ±10.25	94.75% ±0.17	88.53% ±1.11	63.47% ±2.57	94.44% ±1.28	88.74% ±1.70	47.94% ±24.66
65	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.017	m1	96.51% ±0.07	89.48% ±0.59	67.81% ±1.08	95.83% ±0.26	88.59% ±1.50	70.12% ±3.51	95.69% ±0.12	88.73% ±0.67	67.54% ±1.39	95.79% ±0.27	88.97% ±1.53	25.26% ±3.18
66	e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1	vinai/phobert-base-v2	0.026	m1	96.76% ±0.21	89.49% ±0.61	70.29% ±4.02	95.38% ±1.01	88.39% ±1.20	72.62% ±5.19	96.07% ±0.20	88.82% ±0.52	70.24% ±3.80	95.34% ±1.02	88.77% ±1.22	38.89% ±14.70
67	e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.065	m1	96.42% ±0.15	88.58% ±0.73	67.34% ±2.00	94.24% ±1.00	89.27% ±0.84	73.00% ±5.87	95.61% ±0.22	87.78% ±0.71	66.89% ±2.67	94.18% ±1.01	89.45% ±1.18	49.29% ±8.07
68	e2_xlmr_base_cross_logitadjust_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.018	m1	95.80% ±0.32	87.17% ±0.78	62.71% ±2.21	94.19% ±1.22	88.24% ±1.50	69.23% ±0.00	94.64% ±0.40	86.24% ±0.69	62.10% ±2.66	94.13% ±1.23	88.56% ±1.59	42.86% ±0.00
69	e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-large-instruct	0.026	m1	96.55% ±0.48	91.43% ±0.34	71.17% ±4.55	95.02% ±0.43	90.09% ±0.65	73.19% ±1.53	95.69% ±0.65	90.79% ±0.36	70.30% ±4.40	94.97% ±0.43	90.51% ±0.66	52.22% ±1.92
70	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1	jhu-clsp/mmBERT-base	0.072	m1	96.44% ±0.43	89.77% ±0.88	67.38% ±1.92	94.63% ±0.27	86.52% ±4.38	69.39% ±9.12	95.60% ±0.51	88.95% ±0.92	66.66% ±1.71	94.58% ±0.27	86.88% ±4.45	43.32% ±15.43
71	e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1	jhu-clsp/mmBERT-base	0.082	m1	96.27% ±0.05	89.90% ±0.68	67.56% ±1.14	94.71% ±0.88	86.43% ±4.04	68.25% ±7.90	95.38% ±0.15	89.01% ±0.76	67.09% ±1.19	94.65% ±0.89	86.78% ±4.10	43.33% ±14.53
72	e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.067	m1	95.55%	88.30%	60.65%	95.52%	89.53%	74.07%	94.53%	87.44%	60.32%	95.48%	89.94%	53.33%
73	e2_visobert_cross_ce_real_frozenemb_lr4e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	uitnlp/visobert	0.037	m1	94.52%	87.03%	55.13%	92.46%	87.86%	72.00%	93.54%	86.01%	55.21%	92.39%	88.24%	50.00%
74	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1	FacebookAI/xlm-roberta-base	0.027	m1	95.03%	89.62%	61.87%	95.10%	88.51%	72.00%	93.98%	88.69%	61.97%	95.05%	88.89%	46.15%
75	e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m1mmBERT base · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.057	m1	96.10% ±0.31	89.48% ±1.64	65.39% ±3.87	96.43% ±0.47	89.98% ±1.56	73.73% ±3.28	95.34% ±0.26	88.74% ±1.89	65.23% ±3.30	96.39% ±0.48	90.41% ±1.58	39.81% ±13.13
76	e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.206	m1	96.73% ±0.44	89.84% ±1.80	75.09% ±7.76	94.91% ±0.72	89.82% ±1.07	75.33% ±10.98	95.99% ±0.58	89.06% ±1.95	74.59% ±8.19	94.86% ±0.73	90.22% ±1.09	56.59% ±17.96
77	e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.034	m1	96.68% ±0.12	89.19% ±0.41	70.25% ±2.50	95.27% ±0.55	89.35% ±0.52	73.81% ±6.30	95.85% ±0.21	88.36% ±0.52	69.76% ±2.79	95.22% ±0.56	89.31% ±0.72	46.76% ±14.72
78	e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.052	m1	96.77% ±0.48	88.91% ±0.79	69.35% ±4.22	95.87% ±1.12	87.50% ±3.31	69.00% ±6.46	95.98% ±0.57	88.07% ±0.84	68.59% ±4.03	95.83% ±1.13	87.87% ±3.37	36.19% ±6.60
79	e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.033	m1	96.62% ±0.28	88.80% ±1.06	69.16% ±2.96	95.06% ±1.24	87.22% ±3.11	70.69% ±8.52	95.88% ±0.36	88.06% ±1.16	68.89% ±2.74	95.01% ±1.25	87.59% ±3.15	37.54% ±17.00
80	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1	jhu-clsp/mmBERT-base	0.144	m1	96.01% ±0.23	89.78% ±0.48	65.85% ±2.92	94.02% ±0.49	88.33% ±1.00	69.11% ±2.38	95.03% ±0.28	88.96% ±0.47	64.84% ±2.83	93.96% ±0.50	88.88% ±1.13	42.06% ±8.36
81	e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.082	m1	96.12% ±0.71	89.37% ±0.69	65.76% ±4.44	94.95% ±0.96	87.58% ±1.63	76.08% ±11.13	95.28% ±0.89	88.56% ±0.71	65.23% ±4.34	94.90% ±0.97	87.95% ±1.66	61.52% ±16.17
82	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1	jhu-clsp/mmBERT-base	0.108	m1	96.66% ±0.14	89.96% ±0.97	70.83% ±3.29	95.30% ±1.20	86.92% ±1.50	72.53% ±3.96	95.88% ±0.17	89.20% ±0.99	70.28% ±3.19	95.25% ±1.21	87.49% ±1.88	50.00% ±7.14
83	e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-0.6B-Base	0.075	m1	96.13% ±0.54	88.87% ±1.26	65.39% ±5.94	94.61% ±0.93	92.36% ±2.02	72.95% ±1.36	95.21% ±0.75	87.89% ±1.38	64.76% ±6.22	94.55% ±0.94	92.81% ±2.05	54.07% ±1.28
84	e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1	Qwen/Qwen3-0.6B-Base	0.077	m1	94.82% ±0.32	87.70% ±1.96	60.26% ±3.59	95.61% ±1.06	90.93% ±1.58	75.64% ±8.41	94.08% ±0.89	86.70% ±2.13	61.81% ±5.58	95.56% ±1.07	91.36% ±1.61	52.38% ±17.17
85	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_m1	vinai/phobert-base-v2	0.093	m1	96.44% ±0.67	90.12% ±1.06	69.63% ±1.01	94.99% ±1.46	88.56% ±1.40	77.28% ±1.98	95.61% ±0.86	89.44% ±1.12	68.91% ±0.97	94.94% ±1.47	88.94% ±1.43	59.02% ±4.07
86	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m1	vinai/phobert-base-v2	0.104	m1	96.65% ±0.35	89.34% ±0.91	71.31% ±6.00	95.10% ±1.28	87.70% ±2.27	78.02% ±3.42	95.82% ±0.45	88.56% ±0.85	70.68% ±6.05	95.05% ±1.29	88.07% ±2.31	54.96% ±5.94
87	e2_xlmr_base_cross_ce_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.044	m1	95.85% ±0.34	87.99% ±2.24	64.36% ±5.36	93.16% ±0.53	88.77% ±0.37	69.09% ±5.23	94.88% ±0.41	87.13% ±2.68	64.09% ±5.85	93.09% ±0.53	89.36% ±0.73	36.57% ±16.67
88	e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.019	m1	95.84% ±0.26	87.81% ±0.65	66.15% ±3.68	94.00% ±0.25	88.08% ±0.70	74.00% ±3.97	94.77% ±0.35	86.79% ±0.77	65.64% ±4.10	93.95% ±0.25	88.45% ±0.70	52.88% ±8.55
89	e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.049	m1	96.35% ±0.19	88.38% ±0.85	66.55% ±0.38	95.19% ±0.21	88.76% ±1.47	68.25% ±2.75	95.52% ±0.18	87.60% ±1.01	66.22% ±0.67	95.32% ±0.52	89.14% ±1.49	42.59% ±8.49
90	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.159	m1	96.40% ±0.20	88.30% ±1.28	69.27% ±6.18	95.25% ±1.02	88.15% ±2.17	61.51% ±23.33	95.61% ±0.34	87.56% ±1.34	69.21% ±5.57	95.21% ±1.03	88.53% ±2.20	48.13% ±11.70
91	e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-4B	0.061	m1	95.09% ±0.09	86.16% ±1.21	62.48% ±1.68	93.19% ±5.21	84.26% ±0.44	71.46% ±14.64	93.74% ±0.19	85.02% ±1.25	61.23% ±2.14	93.11% ±5.27	84.24% ±0.05	50.83% ±12.96
92	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.025	m1	96.66%	88.78%	68.04%	94.12%	85.06%	25.00%	95.85%	88.12%	68.97%	94.06%	85.38%	22.22%
93	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.152	m1	95.43% ±0.35	87.18% ±1.41	64.46% ±3.74	92.62% ±0.71	84.44% ±1.67	69.76% ±4.64	94.22% ±0.47	86.06% ±1.57	63.18% ±4.13	92.55% ±0.72	84.76% ±1.69	49.22% ±6.76
94	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m1	vinai/phobert-base-v2	0.035	m1	96.35% ±0.51	89.07% ±1.18	68.00% ±4.10	94.91% ±1.03	89.06% ±1.06	73.81% ±2.06	95.51% ±0.59	88.23% ±1.38	67.29% ±4.10	94.86% ±1.04	89.45% ±1.08	50.00% ±0.00
95	e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-base	0.052	m1	95.53%	88.91%	62.57%	91.79%	88.76%	33.33%	94.42%	88.02%	62.68%	91.71%	89.16%	36.36%
96	e2_visobert_cross_ce_real_frozenemb_lr2e5_m1	uitnlp/visobert	0.027	m1	94.79%	85.30%	57.07%	93.94%	88.64%	58.33%	93.76%	84.17%	57.07%	93.88%	89.02%	16.67%
97	e2_visobert_cross_ce_real_frozenemb_lr1e5_m1	uitnlp/visobert	0.186	m1	94.60%	85.71%	58.33%	93.47%	82.14%	61.54%	93.29%	84.82%	57.85%	93.40%	82.21%	28.57%
98	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1	FacebookAI/xlm-roberta-base	0.161	m1	95.25%	88.15%	66.48%	92.31%	86.55%	51.85%	94.01%	87.42%	66.27%	92.23%	86.90%	13.33%
99	e1_tfidf_hybrid_qwen3emb_lr_m1TF-IDF câu KH (char 1-5, word 1-2-gram, giữ từ 1 ký tự, thêm kênh char bỏ dấu) + câu bot (char 2-4) + tương tác loại câu bot × từ câu KH · vector đóng băng Qwen3-Embedding-0.6B (cặp bot + KH, pooling last_token), trọng số dense 1 · LogReg · train thật, nhãn m1	Qwen/Qwen3-Embedding-0.6B	0.149	m1	95.31%	84.20%	60.27%	91.19%	85.54%	69.57%	94.33%	82.81%	60.06%	91.10%	85.89%	36.36%
100	e1_tfidf_svmcal_ce_synth1x_fbmatched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	TF-IDF	0.207	m1	95.57%	86.96%	67.52%	93.00%	86.06%	72.73%	94.49%	85.90%	66.89%	92.93%	86.42%	40.00%
101	e1_tfidf_svmcal_ce_synth1x_nofb_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent omi_intent_repeat, callback, relationship, negate, affirm), nhãn m1	TF-IDF	0.164	m1	95.63%	86.90%	65.42%	93.00%	86.06%	69.57%	94.56%	85.84%	64.69%	92.93%	86.42%	36.36%
102	e1_tfidf_svmcal_ce_synth1x_noqparticle_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm q_particle), nhãn m1	TF-IDF	0.280	m1	94.94%	85.97%	60.27%	92.54%	85.89%	80.00%	93.83%	84.79%	60.06%	92.46%	86.25%	61.54%
103	e1_tfidf_svmcal_ce_real_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn m1	TF-IDF	0.178	m1	95.70%	86.78%	66.26%	93.00%	86.06%	69.57%	94.65%	85.59%	65.81%	92.93%	86.42%	36.36%
104	e1_tfidf_svmcal_ce_synth1x_nofiller_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm filler_only), nhãn m1	TF-IDF	0.267	m1	95.08%	85.74%	61.13%	92.54%	85.19%	78.26%	93.78%	84.53%	60.11%	92.46%	85.53%	54.55%
105	e1_tfidf_svmcal_ce_synth1x_noofftopic_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm prefix_offtopic), nhãn m1	TF-IDF	0.298	m1	94.95%	86.17%	61.45%	92.54%	86.06%	76.92%	93.93%	85.02%	61.68%	92.46%	86.42%	57.14%
106	e2_dual_qwen3emb06b_concat_ce_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.227	m1	93.87% ±0.53	83.06% ±0.56	59.33% ±2.53	93.41% ±1.42	76.01% ±4.00	77.23% ±4.17	92.38% ±0.65	81.77% ±0.49	58.54% ±2.56	93.34% ±1.44	76.18% ±4.07	44.34% ±15.71
107	e1_tfidf_svmcal_ce_synth1x_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1, nhãn m1	TF-IDF	0.282	m1	94.91%	85.71%	59.89%	92.54%	85.37%	76.92%	93.74%	84.50%	59.48%	92.46%	85.71%	57.14%
108	e1_tfidf_svmcal_ce_synth1x_matched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ intent của train thật), nhãn m1	TF-IDF	0.190	m1	95.67%	86.69%	66.26%	93.00%	85.37%	69.57%	94.57%	85.50%	65.58%	92.93%	85.71%	36.36%
109	e1_tfidf_svmcal_ce_synth1x_nomoney_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm money), nhãn m1	TF-IDF	0.310	m1	94.94%	86.20%	62.01%	92.54%	86.75%	80.00%	93.74%	85.05%	61.54%	92.46%	87.12%	61.54%
110	e1_tfidf_svmcal_ce_synth1x_fbonly_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent nlu_fallback), nhãn m1	TF-IDF	0.356	m1	94.36%	86.09%	57.96%	92.54%	86.06%	76.92%	93.30%	84.79%	58.59%	92.46%	86.42%	57.14%
–	e2_bamibert_cross_ce_real_frozenemb_lr1e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_bamibert_cross_ce_real_frozenemb_lr2e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_bamibert_cross_ce_real_frozenemb_lr4e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e3_fbverifier_devtop1_familygeomean_fbf1_m1	–	hard	m1	97.91%	93.00%	84.41%	95.61%	90.91%	75.86%	97.54%	92.67%	84.68%	95.57%	91.33%	58.82%
–	e3_fbverifier_devtop1_familygeomean_macrof1_m1	–	hard	m1	97.91%	93.00%	84.41%	95.61%	90.91%	75.86%	97.54%	92.67%	84.68%	95.57%	91.33%	58.82%
1	e3_ensemble_mean_all	–	0.065	v1	95.21%	88.07%	63.74%	91.78%	85.71%	54.55%	93.84%	86.94%	62.45%	92.31%	86.13%	44.44%
2	e3_ensemble_greedy_all	–	0.070	v1	95.06%	88.34%	63.47%	91.89%	84.89%	54.55%	93.71%	87.25%	62.40%	92.41%	85.29%	44.44%
3	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5	–	0.023	v1	95.20%	89.10%	66.12%	89.93%	82.61%	37.50%	93.91%	88.41%	64.60%	90.41%	82.96%	28.57%
4	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1	vinai/phobert-base-v2	0.011	v1	95.04%	85.87%	63.53%	93.79%	86.52%	40.00%	93.84%	84.64%	63.83%	94.37%	86.96%	25.00%
5	e2_xlmr_base_cross_wcesqrt_real_frozenemb	FacebookAI/xlm-roberta-base	0.021	v1	94.87% ±0.10	88.39% ±0.54	65.20% ±1.74	92.36% ±1.27	85.05% ±1.42	53.03% ±18.37	93.42% ±0.24	87.44% ±0.42	64.39% ±1.12	92.89% ±1.30	85.44% ±1.45	42.96% ±22.37
6	e3_prior_em_mmbert_bottype	–	0.082	v1	95.08%	73.97%	69.53%	89.36%	42.64%	57.14%	93.87%	69.93%	68.87%	89.86%	41.06%	40.00%
7	e3_prior_em_mmbert_global	–	0.082	v1	94.83%	86.88%	65.37%	91.67%	82.09%	57.14%	93.45%	86.12%	63.93%	92.20%	82.44%	40.00%
8	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5	jhu-clsp/mmBERT-base	0.061	v1	95.32% ±0.14	88.07% ±1.30	63.86% ±2.84	91.05% ±1.67	85.88% ±2.86	74.24% ±8.44	94.18% ±0.22	87.17% ±1.50	63.20% ±3.11	91.57% ±1.71	86.29% ±2.92	68.89% ±10.18
9	e2_xlmr_base_cross_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	95.14% ±0.17	88.61% ±0.76	67.48% ±0.98	92.06% ±0.40	84.15% ±1.07	55.79% ±10.31	93.85% ±0.17	87.74% ±0.85	66.95% ±1.33	92.59% ±0.40	84.53% ±1.09	45.98% ±10.47
10	e2_xlmr_base_cross_logitadjust_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	95.04% ±0.44	88.03% ±0.40	63.79% ±3.19	92.73% ±1.69	86.46% ±1.90	46.34% ±3.57	93.65% ±0.53	87.09% ±0.46	62.82% ±2.50	93.28% ±1.73	86.89% ±1.94	37.52% ±4.87
11	e2_qwen3_06b_base_cross_ce_real_lora	Qwen/Qwen3-0.6B-Base	0.098	v1	94.55%	86.82%	66.67%	92.41%	84.51%	54.55%	93.14%	85.71%	66.38%	92.96%	84.89%	44.44%
12	e3_ensemble_stack_all	–	0.088	v1	94.38%	89.01%	64.34%	92.20%	84.67%	60.00%	92.70%	88.04%	64.14%	92.75%	85.07%	66.67%
13	e2_qwen3_17b_base_cross_ce_real_lora	Qwen/Qwen3-1.7B-Base	0.083	v1	94.96% ±0.09	87.79% ±0.35	62.72% ±1.68	92.81% ±0.35	86.86% ±1.64	69.46% ±12.92	93.67% ±0.08	86.84% ±0.45	62.11% ±1.90	93.37% ±0.36	87.30% ±1.67	63.30% ±16.33
14	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5	vinai/phobert-base-v2	0.022	v1	95.22% ±0.16	88.54% ±0.67	66.58% ±0.83	92.71% ±1.27	84.91% ±2.46	58.33% ±14.43	93.96% ±0.24	87.59% ±0.75	65.99% ±1.11	93.26% ±1.31	85.30% ±2.52	44.44% ±19.25
15	e2_me5_base_cross_ce_real_frozenemb_lr1e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	intfloat/multilingual-e5-base	0.088	v1	94.65%	87.72%	66.67%	89.19%	84.89%	50.00%	93.01%	86.71%	64.41%	89.66%	85.29%	40.00%
16	e2_me5_base_cross_ce_real_frozenemb_lr4e5	–	0.028	v1	95.19%	88.38%	67.47%	92.52%	83.69%	66.67%	93.90%	87.44%	66.38%	93.06%	84.06%	57.14%
17	e2_xlmr_base_khonly_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.076	v1	94.59% ±0.34	87.90% ±1.43	66.84% ±2.57	91.68% ±0.81	86.38% ±1.62	55.72% ±14.99	93.13% ±0.50	86.90% ±1.65	66.16% ±2.59	92.20% ±0.83	86.81% ±1.65	42.86% ±20.76
18	e2_me5_base_cross_ce_real_frozenemb_lr2e5	intfloat/multilingual-e5-base	0.056	v1	95.02% ±0.23	88.98% ±0.34	66.67% ±1.75	91.07% ±1.02	85.04% ±1.76	50.37% ±14.29	93.65% ±0.25	88.25% ±0.41	65.50% ±1.81	91.57% ±1.04	85.45% ±1.80	36.90% ±17.62
19	e2_gemma3_270m_cross_ce_real_lora	unsloth/gemma-3-270m	0.116	v1	94.61% ±0.43	87.68% ±0.53	62.98% ±2.10	90.62% ±0.95	82.40% ±1.26	51.52% ±7.61	93.19% ±0.60	86.47% ±0.59	62.67% ±2.82	91.34% ±0.98	82.95% ±1.01	42.59% ±8.49
20	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen	Qwen/Qwen3-Embedding-0.6B	0.114	v1	94.29% ±0.14	86.13% ±0.49	61.02% ±1.17	88.70% ±0.83	87.42% ±2.00	62.22% ±7.70	92.78% ±0.31	84.91% ±0.56	59.98% ±0.90	89.15% ±0.86	87.87% ±2.05	55.38% ±7.99
21	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5	–	0.075	v1	95.29%	88.40%	63.12%	92.31%	85.71%	54.55%	94.31%	87.63%	63.64%	92.86%	86.13%	44.44%
22	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5	–	0.026	v1	94.81%	86.85%	65.87%	91.16%	84.29%	57.14%	93.44%	86.11%	65.22%	91.67%	84.67%	66.67%
23	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	FacebookAI/xlm-roberta-base	0.123	v1	94.45%	89.54%	65.52%	91.50%	82.96%	36.36%	93.03%	88.79%	65.09%	92.00%	83.33%	22.22%
24	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5	jhu-clsp/mmBERT-base	0.120	v1	94.87% ±0.16	88.28% ±1.08	65.42% ±0.24	93.43% ±1.44	85.49% ±0.85	45.00% ±7.07	93.77% ±0.06	87.51% ±1.20	65.62% ±0.17	94.00% ±1.47	86.18% ±1.26	35.38% ±6.53
25	e2_visobert_cross_ce_real_frozenemb_lr2e5	–	0.093	v1	94.50%	87.62%	62.60%	90.91%	81.69%	46.15%	93.15%	86.90%	62.56%	91.43%	82.01%	36.36%
26	e2_visobert_cross_ce_real_frozenemb_lr4e5	–	0.066	v1	94.18%	87.72%	60.31%	91.67%	82.86%	66.67%	92.57%	86.88%	59.75%	92.20%	84.06%	60.00%
27	e1_tfidf_svmcal_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn v1	TF-IDF	0.123	v1	94.74%	86.38%	63.88%	90.41%	84.21%	66.67%	93.35%	85.06%	63.07%	90.91%	84.62%	57.14%
28	e1_tfidf_logreg_balanced_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg cân bằng lớp · train thật, nhãn v1	TF-IDF	0.256	v1	94.13%	82.95%	59.26%	87.84%	86.30%	53.33%	92.37%	81.34%	57.60%	88.28%	85.92%	46.15%
29	e1_tfidf_logreg_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật, nhãn v1	TF-IDF	0.125	v1	94.32%	84.53%	60.14%	88.89%	81.48%	60.00%	92.62%	82.97%	58.59%	89.36%	81.82%	50.00%
30	e2_visobert_cross_ce_real_frozenemb_lr1e5	–	0.232	v1	94.29%	86.75%	60.34%	89.19%	80.85%	42.86%	92.87%	86.10%	59.43%	89.66%	81.16%	33.33%
31	e1_tfidf_logreg_ce_synth1x_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật + synth ×1, nhãn v1	TF-IDF	0.234	v1	94.15%	84.34%	60.47%	89.66%	81.48%	54.55%	92.77%	82.77%	60.43%	90.14%	81.82%	44.44%
–	e2_mdeberta_cross_ce_real_frozenemb_lr1e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	0.00%	38.47%	0.00%	0.00%	0.00%	0.00%	0.00%	42.06%	0.00%	0.00%	0.00%	0.00%
–	e2_videberta_cross_ce_real_frozenemb_lr1e5	–	argmax	v1	0.00%	38.47%	0.00%	0.00%	0.00%	0.00%	0.00%	42.06%	0.00%	0.00%	0.00%	0.00%
–	e2_videberta_cross_ce_real_frozenemb_lr2e5	–	argmax	v1	0.00%	38.47%	0.00%	0.00%	0.00%	0.00%	0.00%	42.06%	0.00%	0.00%	0.00%	0.00%
–	e2_videberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_videberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	0.00%	38.47%	0.00%	0.00%	0.00%	0.00%	0.00%	42.06%	0.00%	0.00%	0.00%	0.00%
1	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_lv21	jhu-clsp/mmBERT-base	0.076	v2_1	94.92% ±0.36	88.37% ±0.83	64.36% ±3.05	92.42% ±0.76	86.37% ±3.04	60.85% ±14.40	93.78% ±0.37	87.95% ±0.76	64.92% ±2.62	92.98% ±0.78	86.80% ±3.09	50.79% ±19.83
2	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_lv21	vinai/phobert-base-v2	0.033	v2_1	95.16%	88.45%	67.86%	90.14%	79.43%	40.00%	93.79%	87.55%	67.32%	90.65%	79.71%	25.00%

Dev và test_full (dữ liệu liên quan train — không dùng làm số chính)

dev là tập chọn ngưỡng và xếp STT; test_full (full) chứa cả cặp QC trùng train, nên hai tập này lạc quan hơn test thật. Chỉ để đối chiếu; baseline không có dev nên hiện N/A.

STT	Model_Name	Backbone	Threshold	Data	Họ	Máy	Seed	n_eff	Params	Latency ms	dev Accuracy	dev Macro_Pre	dev Macro_Rec	dev Macro_F1	dev Macro_F1 −rep	dev Weighted_F1	dev FB recall	dev FB precision	dev argmax Macro_F1	dev argmax Macro_F1 −rep	dev argmax FB recall	dev argmax FB precision	full Accuracy	full Macro_Pre	full Macro_Rec	full Macro_F1	full Macro_F1 −rep	full Weighted_F1	full FB recall	full FB precision	full argmax Macro_F1	full argmax Macro_F1 −rep	full argmax FB recall	full argmax FB precision	Trạng thái	Phân tích
0	e0_prod_sft_hard_none_qc_v16_m1 tham chiếu · provenance: unknown	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	m1	reference	zephyrus	1	1	2.0 B (reported)	~1,715 ms (reported)	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	94.91%	89.05%	91.37%	89.58%	90.95%	94.84%	69.29%	86.61%	89.58%	90.95%	69.29%	86.61%	ok	–
0	e0_prod_sft_hard_none_qc_v16 tham chiếu · provenance: unknown	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v1	reference	zephyrus	1	1	2.0 B (reported)	~1,715 ms (reported)	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	94.38%	87.07%	89.91%	87.41%	89.90%	94.31%	67.97%	86.67%	87.41%	89.90%	67.97%	86.67%	ok	–
0	e0_prod_sft_hard_none_qc_v16_lv2 tham chiếu · provenance: unknown	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2	reference	zephyrus	1	1	2.0 B (reported)	~1,715 ms (reported)	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	94.34%	88.09%	90.75%	88.83%	90.13%	94.29%	69.01%	81.67%	88.83%	90.13%	69.01%	81.67%	oktest_informed	–
0	e0_prod_sft_hard_none_qc_v16_lv21 tham chiếu · provenance: unknown	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2_1	reference	zephyrus	1	1	2.0 B (reported)	~1,715 ms (reported)	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	94.38%	87.07%	89.91%	87.41%	89.90%	94.31%	67.97%	86.67%	87.41%	89.90%	67.97%	86.67%	ok	–
1	e3_ensemble_devtop1_familygeomean_m1	–	0.051	m1	ensemble	zephyrus	1	1	N/A	271.8 ms	98.40%	98.10%	98.48%	98.24%	98.21%	98.42%	100.00%	90.55%	98.26%	98.38%	96.39%	94.86%	93.32%	86.31%	90.05%	87.02%	88.78%	93.85%	97.86%	53.73%	89.56%	92.30%	86.43%	81.21%	okp85_not_met	–
2	e3_ensemble_devtop1_familygeomean_tempdev_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	ensemble	zephyrus	1	1	N/A	271.8 ms	98.24%	97.94%	98.39%	98.12%	98.06%	98.26%	100.00%	89.89%	98.22%	98.33%	96.39%	94.86%	93.59%	86.58%	90.32%	87.37%	89.19%	94.07%	97.86%	55.24%	89.11%	92.26%	86.43%	81.76%	okp85_not_met	–
3	e3_ensemble_devtop1_familygeomean_temphard_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	ensemble	zephyrus	1	1	N/A	271.8 ms	98.24%	97.94%	98.39%	98.12%	98.06%	98.26%	100.00%	89.89%	98.22%	98.33%	96.39%	94.86%	93.59%	86.58%	90.32%	87.37%	89.19%	94.07%	97.86%	55.24%	89.11%	92.26%	86.43%	81.76%	okp85_not_met	–
4	e3_ensemble_devtop1_familymean_m1	–	0.112	m1	ensemble	zephyrus	1	1	N/A	271.8 ms	98.32%	98.12%	98.44%	98.24%	98.21%	98.34%	100.00%	91.21%	98.22%	98.33%	96.79%	94.88%	93.70%	87.17%	91.79%	88.51%	89.54%	94.14%	97.14%	56.20%	89.37%	92.08%	84.29%	80.82%	okp85_not_met	–
5	e3_ensemble_sweep_k2_m1	–	0.110	m1	ensemble	zephyrus	1	1	N/A	106.3 ms	98.48%	98.34%	98.54%	98.41%	98.41%	98.49%	100.00%	92.57%	97.90%	98.11%	95.18%	94.42%	94.87%	87.71%	91.65%	89.24%	91.08%	95.02%	94.29%	68.04%	89.75%	92.32%	85.00%	82.07%	okp85_not_met	–
6	e3_ensemble_sweep_k3_m1	–	0.051	m1	ensemble	zephyrus	1	1	N/A	45.3 ms	98.36%	97.97%	98.46%	98.16%	98.11%	98.38%	100.00%	89.57%	98.42%	98.42%	96.79%	94.51%	94.01%	84.94%	90.04%	86.68%	89.62%	94.35%	95.71%	59.56%	87.47%	91.17%	85.00%	76.28%	okp85_not_met	–
7	e3_ensemble_sweep_k4_m1	–	0.058	m1	ensemble	zephyrus	1	1	N/A	94.7 ms	98.40%	98.02%	98.48%	98.20%	98.16%	98.42%	100.00%	89.89%	98.58%	98.46%	97.19%	94.53%	94.13%	85.21%	90.23%	86.92%	89.90%	94.47%	95.71%	59.29%	87.72%	91.47%	84.29%	77.12%	okp85_not_met	–
8	e3_ensemble_sweep_k5_m1	–	0.111	m1	ensemble	zephyrus	1	1	N/A	277.3 ms	98.71%	98.51%	98.66%	98.56%	98.59%	98.72%	100.00%	92.91%	98.46%	98.47%	96.79%	94.88%	95.10%	88.70%	92.35%	90.01%	91.35%	95.28%	96.43%	66.83%	90.21%	92.26%	85.71%	81.08%	okp85_not_met	–
9	e3_ensemble_sweep_k6_m1	–	0.087	m1	ensemble	zephyrus	1	1	N/A	277.3 ms	98.71%	98.51%	98.66%	98.56%	98.59%	98.72%	100.00%	92.91%	98.46%	98.47%	96.79%	94.88%	94.79%	89.42%	92.18%	90.12%	90.76%	95.04%	96.43%	63.38%	90.21%	92.26%	85.71%	81.08%	okp85_not_met	–
10	e3_ensemble_sweep_k7_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.087	m1	ensemble	zephyrus	1	1	N/A	277.3 ms	98.71%	98.51%	98.66%	98.56%	98.59%	98.72%	100.00%	92.91%	98.46%	98.47%	96.79%	94.88%	94.79%	89.42%	92.18%	90.12%	90.76%	95.04%	96.43%	63.38%	90.21%	92.26%	85.71%	81.08%	okp85_not_met	–
11	e3_ensemble_sweep_k8_m1	–	0.087	m1	ensemble	zephyrus	1	1	N/A	277.3 ms	98.71%	98.51%	98.66%	98.56%	98.59%	98.72%	100.00%	92.91%	98.46%	98.47%	96.79%	94.88%	94.79%	89.42%	92.18%	90.12%	90.76%	95.04%	96.43%	63.38%	90.21%	92.26%	85.71%	81.08%	okp85_not_met	–
12	e3_ensemble_top_geomean_m1	–	0.073	m1	ensemble	zephyrus	1	1	N/A	149.4 ms	98.44%	98.19%	98.51%	98.31%	98.29%	98.45%	100.00%	91.21%	98.38%	98.37%	96.39%	94.86%	93.94%	85.37%	91.93%	87.77%	89.94%	94.31%	95.71%	58.52%	88.43%	92.37%	86.43%	80.13%	okp85_not_met	–
13	e3_ensemble_top_mean_m1	–	0.073	m1	ensemble	zephyrus	1	1	N/A	149.4 ms	97.97%	97.55%	98.19%	97.79%	97.66%	98.00%	100.00%	87.37%	98.39%	98.38%	96.79%	94.88%	92.85%	84.65%	91.51%	86.75%	88.71%	93.51%	96.43%	50.75%	88.44%	92.38%	84.29%	80.82%	okp85_not_met	–
14	e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	decoder	legion	3	3	24M	94.0 ms	97.93% ±0.31	97.55% ±0.57	98.18% ±0.32	97.80% ±0.44	97.73% ±0.40	97.96% ±0.29	99.60% ±0.00	88.85% ±2.88	98.01% ±0.38	98.08% ±0.20	95.31% ±1.67	95.19% ±0.62	92.40% ±1.44	84.69% ±1.71	87.52% ±0.68	84.76% ±1.04	87.61% ±2.04	93.04% ±1.06	95.00% ±0.71	52.03% ±7.64	87.83% ±0.47	90.71% ±0.49	81.43% ±3.71	78.39% ±2.49	cần sửap85_not_met	–
15	e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1	intfloat/multilingual-e5-base	0.017	m1	encoder	zephyrus	1	1	86M	11.7 ms	97.81%	97.33%	98.05%	97.63%	97.62%	97.84%	99.60%	88.26%	96.62%	97.13%	90.36%	94.14%	91.14%	79.52%	88.11%	82.18%	85.76%	91.96%	94.29%	47.48%	86.26%	89.76%	82.14%	74.68%	okp85_not_met	–
16	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m1	jhu-clsp/mmBERT-base	0.155	m1	encoder	zephyrus	1	1	111M	22.4 ms	98.20%	98.07%	98.31%	98.15%	98.10%	98.22%	99.60%	91.51%	97.23%	97.45%	92.37%	93.12%	93.28%	83.65%	88.64%	85.51%	87.79%	93.52%	87.14%	60.40%	85.90%	89.33%	80.00%	70.89%	okp85_not_met	–
17	e2_qwen3_17b_base_genlabel_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.081	m1	decoder	legion	1	1	17M	55.4 ms	97.42%	97.05%	97.41%	97.13%	97.04%	97.47%	99.60%	85.52%	97.23%	97.30%	93.98%	92.13%	92.58%	82.78%	91.70%	85.92%	87.39%	93.21%	91.43%	50.79%	85.90%	89.75%	79.29%	70.25%	okp85_not_met	–
18	e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-1.7B-Base	0.105	m1	decoder	legion	3	3	17M	52.8 ms	97.63% ±0.06	97.24% ±0.25	97.76% ±0.27	97.42% ±0.26	97.38% ±0.15	97.67% ±0.05	99.60% ±0.00	87.23% ±0.98	97.82% ±0.11	97.96% ±0.10	97.05% ±0.23	93.31% ±0.74	92.91% ±0.29	85.42% ±1.34	92.43% ±0.31	87.77% ±0.59	88.16% ±0.32	93.51% ±0.26	93.81% ±0.41	51.85% ±0.62	87.50% ±0.77	90.78% ±0.48	85.00% ±1.24	72.41% ±5.17	okp85_not_met	–
19	e3_ensemble_top_greedy_m1	–	0.113	m1	ensemble	zephyrus	1	1	N/A	149.4 ms	98.36%	98.08%	98.38%	98.19%	98.14%	98.38%	99.60%	90.51%	98.21%	98.32%	95.98%	94.84%	93.59%	85.49%	92.89%	88.06%	89.37%	94.08%	95.00%	54.96%	89.43%	92.65%	86.43%	81.76%	okp85_not_met	–
20	e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m1	intfloat/multilingual-e5-large-instruct	0.017	m1	encoder	legion	3	3	304M	12.3 ms	97.98% ±0.37	97.48% ±0.53	98.00% ±0.71	97.67% ±0.63	97.73% ±0.42	98.01% ±0.36	99.46% ±0.23	88.30% ±2.78	97.81% ±0.56	97.89% ±0.30	94.78% ±1.39	93.78% ±0.19	93.12% ±0.86	84.26% ±0.42	90.50% ±0.71	86.31% ±0.29	88.51% ±1.47	93.58% ±0.65	94.76% ±0.41	56.57% ±7.19	86.42% ±0.79	90.62% ±0.13	83.57% ±1.24	75.04% ±2.14	okp85_not_met	–
21	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1	jhu-clsp/mmBERT-base	0.108	m1	encoder	zephyrus	3	3	111M	17.0 ms	97.58% ±0.07	97.12% ±0.13	97.71% ±0.06	97.32% ±0.10	97.20% ±0.08	97.63% ±0.06	99.46% ±0.23	85.60% ±0.46	97.85% ±0.30	97.73% ±0.21	95.45% ±1.23	92.28% ±1.90	90.91% ±1.31	82.30% ±0.53	88.88% ±2.55	83.96% ±1.94	85.24% ±1.45	91.81% ±0.92	93.10% ±5.17	46.49% ±6.91	86.43% ±0.92	88.77% ±2.23	87.38% ±5.95	67.34% ±12.16	oktest_informedp85_not_met	–
22	e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1	Qwen/Qwen3-4B-Base	0.056	m1	decoder	legion	3	3	33M	138.1 ms	97.91% ±0.13	97.41% ±0.12	98.09% ±0.29	97.67% ±0.18	97.57% ±0.17	97.94% ±0.12	99.46% ±0.23	86.91% ±0.75	98.22% ±0.12	98.23% ±0.14	96.52% ±0.23	94.25% ±0.94	93.24% ±0.37	83.53% ±0.31	91.10% ±0.72	86.13% ±0.14	89.04% ±0.40	93.74% ±0.30	95.71% ±0.71	54.24% ±2.06	87.95% ±0.21	91.61% ±0.54	81.43% ±3.11	78.01% ±3.63	okp85_not_met	–
23	e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1	unsloth/gemma-3-1b-pt	0.045	m1	decoder	legion	2 1 lỗi	2	13M	51.1 ms	97.70% ±0.28	97.31% ±0.29	97.89% ±0.35	97.51% ±0.33	97.41% ±0.29	97.73% ±0.27	99.40% ±0.28	86.40% ±1.31	98.14% ±0.25	98.15% ±0.19	95.78% ±0.85	94.83% ±1.11	92.54% ±0.16	85.24% ±2.68	88.95% ±1.33	85.83% ±0.52	87.61% ±0.77	93.17% ±0.12	95.36% ±0.51	50.97% ±1.37	87.51% ±0.04	90.73% ±0.70	85.71% ±2.02	73.84% ±0.13	cần sửap85_not_met	–
24	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepeval_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1, nhãn m1	vinai/phobert-base-v2	0.033	m1	encoder	zephyrus	3	3	86M	19.2 ms	97.63% ±0.05	97.18% ±0.07	97.89% ±0.06	97.44% ±0.07	97.25% ±0.08	97.67% ±0.04	99.33% ±0.23	85.98% ±0.46	97.85% ±0.18	97.74% ±0.22	94.78% ±1.61	94.54% ±1.00	91.88% ±0.34	82.95% ±1.72	87.17% ±0.90	83.83% ±1.28	86.69% ±0.09	92.55% ±0.22	92.38% ±1.49	49.21% ±2.29	87.95% ±1.26	90.86% ±0.72	84.52% ±1.09	76.19% ±0.85	cần sửatest_informedp85_not_met	–
25	e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1	unsloth/gemma-3-1b-pt	0.045	m1	decoder	legion	3	3	26M	49.5 ms	97.75% ±0.31	97.22% ±0.48	98.09% ±0.34	97.57% ±0.42	97.40% ±0.35	97.79% ±0.30	99.33% ±0.61	86.20% ±1.82	98.19% ±0.30	98.19% ±0.22	95.18% ±1.06	95.31% ±0.80	91.76% ±1.10	82.93% ±2.16	89.53% ±1.96	84.84% ±2.22	86.88% ±1.68	92.51% ±0.80	94.29% ±1.24	48.95% ±5.81	85.66% ±0.70	90.14% ±0.22	81.19% ±2.70	76.31% ±1.27	okp85_not_met	–
26	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1	jhu-clsp/mmBERT-base	0.052	m1	encoder	zephyrus	3	3	111M	18.0 ms	97.92% ±0.33	97.33% ±0.33	98.16% ±0.29	97.68% ±0.32	97.58% ±0.38	97.95% ±0.32	99.33% ±0.23	87.51% ±1.41	98.02% ±0.18	97.99% ±0.16	95.31% ±0.93	94.07% ±1.22	92.14% ±0.65	81.34% ±1.02	90.47% ±1.27	84.17% ±0.92	86.72% ±0.79	92.73% ±0.52	91.67% ±1.49	51.97% ±3.78	85.70% ±0.48	90.03% ±0.56	82.14% ±4.95	74.42% ±3.45	okp85_not_met	–
27	e2_mmbert_base_cross_ce_synth1x_frozenemb_lr2e5_fbmatched_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	jhu-clsp/mmBERT-base	0.051	m1	encoder	zephyrus	3	3	111M	18.6 ms	98.05% ±0.24	97.65% ±0.26	98.22% ±0.33	97.88% ±0.25	97.77% ±0.30	98.07% ±0.23	99.33% ±0.61	88.97% ±1.09	97.97% ±0.05	97.83% ±0.13	93.31% ±1.41	94.70% ±0.36	92.90% ±0.68	84.17% ±1.08	90.26% ±0.73	86.13% ±1.18	87.86% ±0.67	93.42% ±0.52	92.38% ±2.70	53.74% ±4.22	86.67% ±1.29	89.91% ±0.55	79.05% ±2.89	73.46% ±1.63	oktest_informedp85_not_met	–
28	e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.118	m1	encoder	zephyrus	3	3	111M	18.7 ms	97.91% ±0.08	97.60% ±0.19	98.11% ±0.16	97.79% ±0.10	97.72% ±0.17	97.93% ±0.08	99.33% ±0.46	88.56% ±1.42	98.01% ±0.09	97.98% ±0.07	95.72% ±1.67	93.61% ±1.06	92.52% ±0.92	83.54% ±1.99	89.72% ±1.52	85.45% ±1.95	87.34% ±1.06	93.10% ±0.66	93.33% ±3.52	52.67% ±6.69	86.96% ±0.17	89.93% ±0.70	83.57% ±1.24	71.98% ±5.54	okp85_not_met	–
29	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1	vinai/phobert-base-v2	0.008	m1	encoder	zephyrus	1	1	86M	12.0 ms	98.05%	97.74%	97.00%	97.27%	97.88%	98.07%	99.20%	89.17%	98.13%	98.38%	95.98%	96.37%	92.58%	87.84%	79.05%	77.54%	87.34%	92.94%	93.57%	53.91%	85.90%	90.14%	82.86%	78.91%	okp85_not_met	–
30	e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1	Qwen/Qwen3-4B-Base	0.162	m1	decoder	kaggle	1	1	33M	233.2 ms	97.81%	97.46%	98.03%	97.68%	97.54%	97.85%	99.20%	87.90%	97.99%	97.91%	95.18%	94.80%	92.65%	82.65%	94.33%	86.70%	88.04%	93.22%	96.43%	51.92%	88.22%	90.38%	85.00%	70.00%	okp85_not_met	–
31	e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1	Qwen/Qwen3-8B-Base	0.170	m1	decoder	kaggle	1	1	44M	379.1 ms	98.40%	98.09%	98.28%	98.15%	98.10%	98.42%	99.20%	91.14%	97.99%	98.06%	95.98%	94.09%	94.21%	84.57%	93.23%	88.05%	89.94%	94.49%	92.86%	61.61%	88.03%	91.45%	84.29%	76.62%	okp85_not_met	–
32	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_m1	vinai/phobert-base-v2	0.073	m1	encoder	zephyrus	3	3	86M	20.4 ms	97.92% ±0.18	97.49% ±0.41	98.16% ±0.29	97.77% ±0.36	97.64% ±0.22	97.95% ±0.18	99.20% ±0.40	88.67% ±1.86	97.84% ±0.31	97.67% ±0.21	93.71% ±3.01	94.79% ±1.96	92.59% ±0.22	82.95% ±0.76	90.34% ±1.91	85.31% ±0.71	87.70% ±0.64	93.08% ±0.17	95.00% ±0.71	53.37% ±1.20	87.04% ±1.24	90.48% ±0.39	84.76% ±1.65	73.57% ±1.68	oktest_informedp85_not_met	–
33	e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	decoder	legion	3	3	13M	49.8 ms	97.98% ±0.12	97.56% ±0.19	98.20% ±0.06	97.81% ±0.09	97.64% ±0.19	98.02% ±0.11	99.20% ±0.70	87.50% ±1.33	98.21% ±0.12	98.11% ±0.06	95.05% ±0.46	94.42% ±0.38	92.81% ±0.07	83.55% ±0.69	88.84% ±1.31	85.21% ±0.88	88.03% ±0.52	93.28% ±0.16	93.57% ±2.58	54.56% ±1.46	86.57% ±1.46	90.63% ±0.75	85.24% ±2.89	75.84% ±1.91	okp85_not_met	–
34	e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-base	0.031	m1	encoder	zephyrus	3	3	86M	9.0 ms	97.40% ±0.30	97.11% ±0.39	97.67% ±0.10	97.30% ±0.26	97.19% ±0.38	97.44% ±0.29	99.20% ±0.40	86.29% ±1.98	97.00% ±0.17	97.02% ±0.30	92.24% ±1.62	92.62% ±0.80	92.27% ±0.88	82.70% ±0.15	87.14% ±0.49	83.75% ±0.42	87.16% ±1.36	92.85% ±0.68	95.48% ±1.65	52.58% ±5.12	84.42% ±0.49	88.89% ±1.07	82.86% ±3.11	68.26% ±6.20	okp85_not_met	–
35	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.061	m1	encoder	zephyrus	3	3	111M	17.6 ms	97.84% ±0.43	97.46% ±0.48	98.05% ±0.33	97.69% ±0.40	97.60% ±0.45	97.87% ±0.43	99.20% ±0.80	88.13% ±1.98	97.74% ±0.18	97.80% ±0.04	93.98% ±1.06	94.36% ±1.06	92.07% ±0.86	83.25% ±1.21	88.62% ±0.69	84.66% ±0.93	87.02% ±1.04	92.72% ±0.63	94.05% ±1.09	50.88% ±5.44	86.18% ±1.61	89.92% ±0.90	80.71% ±3.27	73.60% ±5.21	okp85_not_met	–
36	e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1	Qwen/Qwen3-4B-Base	0.131	m1	decoder	kaggle	3	3	33M	232.8 ms	97.83% ±0.46	97.45% ±0.58	98.06% ±0.27	97.69% ±0.46	97.54% ±0.55	97.86% ±0.44	99.20% ±0.70	87.58% ±3.37	98.18% ±0.14	98.18% ±0.08	96.79% ±0.40	94.02% ±0.57	93.83% ±0.84	84.97% ±2.36	91.98% ±0.97	87.35% ±0.78	89.68% ±1.08	94.20% ±0.63	94.76% ±2.51	60.01% ±8.76	88.21% ±0.79	91.76% ±0.46	83.81% ±2.70	79.45% ±0.50	okp85_not_met	–
37	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.043	m1	encoder	zephyrus	3	3	86M	20.9 ms	97.40% ±0.38	97.14% ±0.38	97.28% ±0.34	97.12% ±0.39	97.19% ±0.37	97.43% ±0.37	99.06% ±0.61	86.69% ±2.16	97.43% ±0.68	97.43% ±0.57	94.24% ±2.82	93.51% ±0.69	92.28% ±0.43	84.55% ±3.02	88.30% ±2.08	85.16% ±1.22	87.48% ±0.78	92.83% ±0.34	92.38% ±2.06	52.04% ±2.31	86.88% ±0.37	89.99% ±0.15	85.24% ±0.82	70.35% ±1.45	okp85_not_met	–
38	e2_gemma3_1b_pt_cross_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.053	m1	decoder	legion	3	3	13M	49.8 ms	97.71% ±0.06	97.12% ±0.25	97.83% ±0.23	97.39% ±0.03	97.39% ±0.12	97.75% ±0.05	99.06% ±0.23	86.25% ±0.75	97.82% ±0.30	97.85% ±0.09	94.11% ±0.61	94.24% ±0.18	92.60% ±0.69	84.04% ±1.46	88.51% ±0.60	85.21% ±0.92	87.61% ±0.97	93.14% ±0.60	94.29% ±0.00	53.12% ±2.91	86.97% ±1.80	90.55% ±0.51	83.10% ±1.80	77.80% ±2.15	okp85_not_met	–
39	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1	vinai/phobert-base-v2	0.037	m1	encoder	zephyrus	3	3	86M	24.1 ms	97.74% ±0.28	97.36% ±0.29	97.81% ±0.61	97.51% ±0.46	97.38% ±0.31	97.77% ±0.28	98.93% ±0.46	87.05% ±0.99	98.05% ±0.24	97.81% ±0.29	94.24% ±1.23	95.67% ±1.24	92.55% ±0.43	85.57% ±1.97	90.27% ±1.48	86.77% ±0.15	87.65% ±0.42	93.05% ±0.33	93.33% ±1.65	53.37% ±2.16	88.87% ±0.79	91.05% ±0.40	83.10% ±2.70	77.72% ±1.83	okp85_not_met	–
40	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m1	vinai/phobert-base-v2	0.085	m1	encoder	zephyrus	3	3	86M	20.7 ms	97.97% ±0.51	97.73% ±0.51	98.15% ±0.39	97.89% ±0.47	97.74% ±0.56	97.99% ±0.50	98.93% ±0.93	89.73% ±2.51	97.96% ±0.24	97.76% ±0.20	95.45% ±0.61	94.20% ±1.10	93.17% ±0.32	84.49% ±1.53	90.09% ±2.19	86.46% ±0.93	88.32% ±0.31	93.50% ±0.30	92.14% ±0.71	58.16% ±1.84	87.37% ±1.28	90.63% ±0.28	83.57% ±1.43	78.00% ±1.16	okp85_not_met	–
41	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_m1	vinai/phobert-base-v2	0.028	m1	encoder	zephyrus	3	3	86M	22.9 ms	97.68% ±0.30	97.28% ±0.29	97.99% ±0.35	97.56% ±0.33	97.33% ±0.31	97.72% ±0.29	98.93% ±0.46	86.64% ±1.02	97.85% ±0.21	97.94% ±0.30	94.78% ±1.39	95.68% ±1.15	92.16% ±0.29	81.71% ±1.76	89.36% ±3.41	84.08% ±2.29	87.20% ±0.08	92.78% ±0.23	93.57% ±1.24	50.66% ±1.25	85.87% ±2.08	91.17% ±0.71	82.14% ±2.47	79.03% ±3.35	okp85_not_met	–
42	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1	jhu-clsp/mmBERT-base	0.082	m1	encoder	zephyrus	3	3	111M	17.8 ms	97.84% ±0.38	97.41% ±0.36	97.93% ±0.27	97.61% ±0.35	97.55% ±0.39	97.87% ±0.36	98.93% ±0.46	87.82% ±2.47	97.66% ±0.24	97.70% ±0.17	93.84% ±0.84	93.49% ±1.44	92.25% ±1.11	81.46% ±1.22	88.61% ±1.41	83.73% ±1.52	86.82% ±1.77	92.80% ±0.85	92.38% ±3.60	53.43% ±7.60	85.69% ±0.64	89.54% ±0.92	82.38% ±4.65	73.47% ±6.59	okp85_not_met	–
43	e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.057	m1	encoder	zephyrus	3	3	111M	19.4 ms	97.71% ±0.12	97.35% ±0.15	97.90% ±0.14	97.56% ±0.16	97.38% ±0.19	97.74% ±0.11	98.93% ±0.84	87.26% ±1.17	97.99% ±0.05	97.85% ±0.07	95.05% ±0.61	93.80% ±1.10	91.90% ±0.30	81.75% ±1.69	90.12% ±0.77	84.24% ±1.06	86.50% ±0.28	92.64% ±0.26	93.10% ±2.18	48.63% ±1.08	85.20% ±1.16	89.50% ±0.58	82.14% ±3.27	70.01% ±4.62	oktest_informedp85_not_met	–
44	e2_qwen3_17b_base_cross_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.101	m1	decoder	legion	3	3	17M	48.0 ms	97.63% ±0.32	97.26% ±0.40	97.66% ±0.31	97.38% ±0.37	97.34% ±0.44	97.67% ±0.30	98.93% ±0.61	87.09% ±2.39	97.62% ±0.63	97.61% ±0.50	93.84% ±2.28	93.33% ±1.27	92.36% ±0.59	82.60% ±1.48	89.91% ±1.96	85.08% ±1.92	87.80% ±1.10	92.94% ±0.41	91.90% ±2.97	51.90% ±4.14	85.76% ±0.52	89.73% ±0.54	79.05% ±0.41	70.05% ±0.82	cần sửap85_not_met	–
45	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m1	vinai/phobert-base-v2	0.052	m1	encoder	zephyrus	3	3	86M	20.9 ms	97.58% ±0.12	97.14% ±0.20	97.40% ±0.39	97.18% ±0.21	97.26% ±0.13	97.62% ±0.12	98.80% ±0.40	86.32% ±0.33	97.74% ±0.25	97.80% ±0.22	95.31% ±0.93	93.74% ±2.49	92.49% ±0.25	84.59% ±1.35	89.11% ±2.40	85.61% ±1.06	87.78% ±0.56	93.03% ±0.14	93.81% ±1.09	52.81% ±3.88	87.44% ±2.06	89.89% ±0.38	82.86% ±4.29	70.77% ±4.52	okp85_not_met	–
46	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1	vinai/phobert-base-v2	0.077	m1	encoder	zephyrus	3	3	86M	21.2 ms	97.55% ±0.23	97.27% ±0.35	97.73% ±0.26	97.43% ±0.34	97.34% ±0.31	97.59% ±0.21	98.80% ±0.70	87.38% ±2.34	97.75% ±0.41	97.61% ±0.24	93.57% ±0.70	95.24% ±1.22	92.95% ±1.09	86.79% ±1.11	89.27% ±1.95	86.78% ±1.38	88.39% ±1.31	93.41% ±0.89	94.52% ±0.41	55.41% ±7.56	89.45% ±0.48	91.61% ±0.03	84.29% ±0.71	80.11% ±1.76	okp85_not_met	–
47	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	vinai/phobert-base-v2	0.058	m1	encoder	zephyrus	3	3	135M	19.6 ms	97.38% ±0.23	97.07% ±0.29	97.41% ±0.55	97.16% ±0.37	97.12% ±0.24	97.42% ±0.23	98.80% ±0.40	86.74% ±1.36	97.39% ±0.15	97.43% ±0.07	94.78% ±1.84	92.95% ±1.57	92.40% ±0.35	82.00% ±1.26	89.68% ±2.49	84.65% ±1.60	87.60% ±0.69	92.88% ±0.29	92.86% ±0.71	54.17% ±2.40	85.96% ±1.23	89.63% ±0.57	85.71% ±0.71	68.22% ±1.82	okp85_not_met	–
48	e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1	unsloth/gemma-3-1b-pt	0.056	m1	decoder	legion	3	3	13M	49.4 ms	97.81% ±0.04	97.34% ±0.12	98.08% ±0.13	97.64% ±0.08	97.48% ±0.10	97.85% ±0.03	98.80% ±0.70	87.03% ±0.96	98.01% ±0.20	98.03% ±0.20	94.91% ±1.41	94.41% ±0.72	92.32% ±0.66	83.36% ±1.40	89.11% ±1.51	84.94% ±0.21	86.98% ±1.25	92.92% ±0.53	93.57% ±0.71	51.31% ±3.86	85.37% ±0.75	89.91% ±0.25	80.00% ±3.11	74.45% ±3.30	okp85_not_met	–
49	e2_gemma3_270m_cross_ce_real_lora_m1gemma-3-270m · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn m1	unsloth/gemma-3-270m	0.098	m1	decoder	zephyrus	3	3	4M	55.8 ms	97.51% ±0.35	97.23% ±0.39	97.93% ±0.51	97.52% ±0.46	97.34% ±0.54	97.54% ±0.34	98.80% ±0.00	88.34% ±2.79	97.72% ±0.41	97.62% ±0.43	93.84% ±1.62	94.60% ±0.88	91.24% ±0.56	80.16% ±0.35	88.04% ±2.12	82.38% ±1.01	85.51% ±1.21	91.89% ±0.49	91.67% ±1.65	49.34% ±2.36	84.18% ±0.84	88.28% ±0.55	79.52% ±1.80	70.54% ±3.55	okp85_not_met	–
50	e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-large-instruct	0.029	m1	encoder	legion	3	3	304M	12.6 ms	97.97% ±0.35	97.66% ±0.35	98.00% ±0.32	97.78% ±0.37	97.76% ±0.46	97.99% ±0.33	98.80% ±1.06	89.53% ±3.05	97.79% ±0.11	98.13% ±0.28	95.18% ±0.00	95.20% ±1.76	93.06% ±0.26	84.47% ±1.11	88.41% ±2.02	85.37% ±0.53	88.20% ±0.61	93.53% ±0.25	93.33% ±1.49	54.55% ±1.30	86.04% ±0.17	90.22% ±0.81	83.10% ±0.41	72.05% ±3.40	cần sửap85_not_met	–
51	e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.034	m1	encoder	zephyrus	3	3	111M	19.7 ms	97.57% ±0.12	97.16% ±0.10	97.71% ±0.30	97.35% ±0.19	97.24% ±0.15	97.61% ±0.12	98.80% ±0.80	86.01% ±0.61	97.78% ±0.11	97.65% ±0.24	93.71% ±1.81	93.71% ±0.45	91.64% ±0.85	82.48% ±2.42	91.64% ±1.05	85.36% ±1.86	86.55% ±0.74	92.43% ±0.70	94.29% ±1.43	48.31% ±2.61	86.55% ±1.28	90.29% ±0.50	81.43% ±3.27	79.20% ±3.81	okp85_not_met	–
52	e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.051	m1	encoder	zephyrus	3	3	111M	18.3 ms	97.59% ±0.06	97.24% ±0.14	97.59% ±0.10	97.34% ±0.13	97.28% ±0.16	97.63% ±0.05	98.80% ±0.40	86.93% ±0.95	97.68% ±0.12	97.64% ±0.23	94.11% ±1.41	93.73% ±0.51	91.98% ±0.53	81.70% ±2.77	89.99% ±3.15	84.22% ±3.24	86.68% ±0.87	92.64% ±0.46	93.57% ±0.71	50.66% ±0.98	85.14% ±2.82	90.06% ±0.67	82.38% ±0.82	74.90% ±0.88	okp85_not_met	–
53	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1	jhu-clsp/mmBERT-base	0.040	m1	encoder	zephyrus	3	3	111M	16.6 ms	97.72% ±0.27	97.28% ±0.23	97.82% ±0.34	97.47% ±0.31	97.45% ±0.21	97.76% ±0.26	98.80% ±1.39	87.16% ±1.42	98.01% ±0.26	97.98% ±0.20	94.91% ±0.61	94.67% ±0.78	91.73% ±1.07	81.75% ±1.68	87.32% ±1.30	83.07% ±1.32	86.30% ±2.04	92.49% ±0.89	94.52% ±0.82	48.49% ±3.91	86.24% ±1.63	90.63% ±0.68	84.05% ±1.80	76.36% ±3.08	okp85_not_met	–
54	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1	jhu-clsp/mmBERT-base	0.057	m1	encoder	zephyrus	3	3	111M	19.6 ms	97.45% ±0.08	96.98% ±0.22	97.68% ±0.23	97.24% ±0.21	97.06% ±0.17	97.49% ±0.08	98.80% ±1.06	85.72% ±0.22	97.61% ±0.27	97.50% ±0.08	92.64% ±0.61	93.90% ±0.37	92.28% ±0.83	82.48% ±1.55	89.73% ±0.36	84.68% ±1.08	87.38% ±0.99	92.90% ±0.64	94.05% ±1.09	51.69% ±4.25	85.89% ±0.68	90.08% ±0.18	79.29% ±0.71	76.75% ±1.52	okp85_not_met	–
55	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.070	m1	encoder	zephyrus	1	1	111M	21.3 ms	98.09%	97.89%	98.10%	97.96%	97.86%	98.11%	98.80%	90.11%	97.87%	97.60%	92.37%	93.88%	93.04%	82.93%	91.14%	85.86%	87.86%	93.44%	92.14%	56.83%	86.56%	89.68%	84.29%	74.68%	okp85_not_met	–
56	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.047	m1	encoder	zephyrus	3	3	111M	20.2 ms	97.93% ±0.20	97.56% ±0.27	98.04% ±0.47	97.75% ±0.27	97.72% ±0.12	97.96% ±0.21	98.80% ±1.45	88.75% ±2.12	97.72% ±0.31	97.67% ±0.30	93.31% ±1.01	93.44% ±0.87	92.20% ±0.22	83.08% ±2.38	87.98% ±0.76	84.40% ±1.28	86.98% ±0.64	92.79% ±0.14	91.90% ±0.82	51.25% ±1.85	85.94% ±1.49	89.59% ±0.63	79.76% ±2.89	75.26% ±1.44	cần sửap85_not_met	–
57	e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1	Qwen/Qwen3-1.7B-Base	0.095	m1	decoder	legion	3	3	17M	52.9 ms	97.74% ±0.20	97.56% ±0.21	97.49% ±0.35	97.47% ±0.27	97.49% ±0.28	97.76% ±0.20	98.80% ±0.80	89.03% ±1.08	97.62% ±0.32	97.77% ±0.41	95.98% ±1.84	93.61% ±0.77	93.28% ±0.70	86.87% ±0.64	90.06% ±2.87	87.47% ±1.54	88.68% ±1.07	93.76% ±0.58	91.90% ±1.80	54.24% ±4.13	88.11% ±0.26	90.87% ±0.37	82.86% ±2.47	73.03% ±2.00	okp85_not_met	–
58	e2_xlmr_base_khonly_ce_real_frozenemb_m1XLM-R base · chỉ câu KH · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1	FacebookAI/xlm-roberta-base	0.012	m1	encoder	zephyrus	1	1	86M	7.9 ms	97.03%	96.91%	97.12%	96.93%	96.96%	97.07%	98.80%	86.32%	97.14%	97.50%	93.98%	94.74%	91.53%	80.58%	87.64%	82.84%	86.61%	92.13%	95.00%	50.96%	85.05%	89.49%	80.71%	78.47%	okp85_not_met	–
59	e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.042	m1	encoder	legion	3	3	100M	55.9 ms	97.53% ±0.06	97.08% ±0.03	97.67% ±0.34	97.29% ±0.18	97.18% ±0.10	97.57% ±0.05	98.66% ±0.61	86.01% ±1.25	97.61% ±0.18	97.65% ±0.16	92.77% ±1.75	94.93% ±0.13	92.05% ±0.33	84.43% ±0.11	88.82% ±1.68	85.36% ±0.98	86.57% ±0.51	92.69% ±0.23	94.29% ±1.89	50.30% ±2.58	87.12% ±0.50	90.27% ±0.57	78.33% ±3.93	79.82% ±3.22	okp85_not_met	–
60	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m1	vinai/phobert-base-v2	0.024	m1	encoder	zephyrus	3	3	86M	19.5 ms	97.45% ±0.13	97.04% ±0.10	97.40% ±0.42	97.13% ±0.27	97.09% ±0.08	97.50% ±0.12	98.66% ±0.23	85.50% ±0.43	97.65% ±0.20	97.54% ±0.44	93.98% ±2.01	93.50% ±2.11	92.06% ±0.36	84.97% ±0.70	88.53% ±2.60	85.35% ±1.18	87.15% ±0.97	92.71% ±0.37	94.52% ±1.09	49.76% ±0.57	87.21% ±0.98	90.57% ±0.65	86.43% ±1.24	72.60% ±0.47	okp85_not_met	–
61	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m1	vinai/phobert-base-v2	0.031	m1	encoder	zephyrus	3	3	86M	22.1 ms	97.65% ±0.16	97.36% ±0.20	97.45% ±0.70	97.33% ±0.33	97.39% ±0.17	97.68% ±0.15	98.66% ±0.61	87.76% ±1.47	97.86% ±0.46	97.74% ±0.57	94.65% ±1.52	94.92% ±2.14	92.37% ±1.10	84.66% ±0.74	88.30% ±2.87	85.26% ±1.91	87.18% ±1.76	92.90% ±0.91	92.62% ±2.06	52.21% ±5.20	86.38% ±1.34	89.61% ±1.38	82.14% ±0.71	70.85% ±5.24	okp85_not_met	–
62	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m1	vinai/phobert-base-v2	0.031	m1	encoder	zephyrus	3	3	135M	22.5 ms	97.41% ±0.16	97.07% ±0.23	97.44% ±0.29	97.17% ±0.24	97.08% ±0.17	97.45% ±0.15	98.66% ±0.46	86.01% ±1.30	97.50% ±0.33	97.47% ±0.38	94.78% ±1.45	92.80% ±1.11	92.56% ±0.73	84.89% ±0.93	88.67% ±1.83	85.69% ±1.40	87.42% ±0.93	93.07% ±0.57	94.52% ±0.41	53.67% ±4.36	86.11% ±0.55	89.68% ±0.52	84.29% ±2.58	71.34% ±2.68	okp85_not_met	–
63	e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1	unsloth/gemma-3-1b-pt	0.096	m1	decoder	legion	3	3	13M	49.3 ms	97.78% ±0.24	97.31% ±0.47	98.10% ±0.39	97.65% ±0.43	97.49% ±0.29	97.81% ±0.24	98.66% ±0.46	87.85% ±1.14	97.79% ±0.24	97.92% ±0.34	94.51% ±1.01	94.39% ±1.57	92.36% ±0.89	79.53% ±3.01	85.24% ±4.98	81.37% ±3.96	87.65% ±1.12	92.87% ±0.68	91.43% ±0.00	53.52% ±5.73	84.48% ±0.83	90.15% ±0.19	82.86% ±0.71	72.25% ±2.42	okp85_not_met	–
64	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1	jhu-clsp/mmBERT-base	0.061	m1	encoder	zephyrus	3	3	111M	19.5 ms	97.48% ±0.33	97.03% ±0.41	97.61% ±0.13	97.23% ±0.16	97.16% ±0.30	97.52% ±0.33	98.66% ±0.84	85.81% ±1.20	97.70% ±0.16	97.61% ±0.04	94.65% ±2.58	92.69% ±1.17	91.16% ±0.52	81.89% ±3.03	89.02% ±4.41	83.43% ±1.44	86.11% ±0.86	92.04% ±0.42	94.29% ±1.89	46.75% ±2.48	85.94% ±1.55	89.64% ±0.29	84.52% ±4.31	70.73% ±2.66	oktest_informedp85_not_met	–
65	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.017	m1	encoder	zephyrus	3	3	86M	14.2 ms	97.94% ±0.16	97.64% ±0.15	97.46% ±0.52	97.49% ±0.34	97.80% ±0.12	97.96% ±0.16	98.66% ±0.61	89.67% ±1.00	97.68% ±0.03	97.99% ±0.14	95.31% ±0.61	94.68% ±0.61	92.82% ±0.23	86.76% ±2.28	86.14% ±0.35	85.00% ±0.90	87.97% ±0.51	93.28% ±0.18	93.10% ±1.65	54.26% ±1.37	85.29% ±1.45	90.10% ±0.69	84.76% ±1.49	71.11% ±2.63	okp85_not_met	–
66	e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1	vinai/phobert-base-v2	0.026	m1	encoder	zephyrus	3	3	135M	9.2 ms	97.98% ±0.06	97.68% ±0.20	97.87% ±0.21	97.74% ±0.12	97.93% ±0.14	98.00% ±0.06	98.66% ±0.23	91.24% ±1.99	97.67% ±0.17	97.97% ±0.09	95.31% ±0.61	94.57% ±1.22	93.16% ±0.66	85.69% ±0.46	87.47% ±1.65	85.70% ±1.41	88.32% ±1.27	93.51% ±0.52	91.43% ±1.24	57.88% ±5.62	86.14% ±0.41	90.18% ±0.44	84.05% ±3.93	72.59% ±2.32	okp85_not_met	–
67	e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.065	m1	encoder	zephyrus	3	3	87M	20.9 ms	97.49% ±0.20	97.04% ±0.29	97.60% ±0.24	97.24% ±0.27	97.17% ±0.17	97.53% ±0.19	98.53% ±0.23	86.29% ±0.83	97.83% ±0.27	97.70% ±0.26	95.85% ±1.16	92.87% ±0.74	92.42% ±0.29	83.17% ±1.71	89.04% ±1.24	85.01% ±1.00	87.41% ±0.61	92.93% ±0.22	93.10% ±1.09	53.30% ±2.55	85.66% ±0.70	89.52% ±0.46	84.52% ±1.80	69.38% ±1.61	okp85_not_met	–
68	e2_xlmr_base_cross_logitadjust_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.018	m1	encoder	zephyrus	3	3	86M	10.6 ms	97.32% ±0.02	96.92% ±0.16	97.55% ±0.26	97.15% ±0.20	97.06% ±0.08	97.36% ±0.02	98.53% ±0.23	86.18% ±0.47	97.21% ±0.05	97.47% ±0.23	93.31% ±0.84	94.06% ±0.24	91.20% ±0.48	80.64% ±0.98	88.50% ±0.49	83.11% ±0.74	85.88% ±0.89	91.95% ±0.38	93.81% ±1.49	47.71% ±2.34	85.15% ±0.36	88.65% ±0.66	84.29% ±2.58	67.28% ±2.45	okp85_not_met	–
69	e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-large-instruct	0.026	m1	encoder	legion	3	3	304M	13.1 ms	98.01% ±0.37	97.70% ±0.56	98.00% ±0.45	97.81% ±0.52	97.79% ±0.44	98.03% ±0.36	98.53% ±0.84	89.67% ±2.24	97.89% ±0.33	98.03% ±0.16	95.05% ±1.62	94.42% ±0.59	93.42% ±0.80	83.99% ±0.70	90.58% ±0.52	86.35% ±0.89	88.98% ±0.99	93.83% ±0.60	92.62% ±1.49	58.15% ±6.44	86.40% ±0.18	90.78% ±0.22	82.38% ±1.80	75.88% ±0.41	okp85_not_met	–
70	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1	jhu-clsp/mmBERT-base	0.072	m1	encoder	zephyrus	3	3	111M	22.0 ms	97.96% ±0.19	97.79% ±0.07	98.04% ±0.46	97.87% ±0.22	97.77% ±0.11	97.98% ±0.20	98.53% ±1.16	89.99% ±0.92	97.72% ±0.32	97.68% ±0.31	93.98% ±2.13	93.23% ±0.74	92.62% ±0.45	82.85% ±2.11	87.41% ±1.39	84.20% ±1.19	87.16% ±0.43	93.06% ±0.30	90.00% ±2.47	54.46% ±3.67	85.83% ±1.58	89.31% ±1.02	80.00% ±2.47	75.49% ±1.15	okp85_not_met	–
71	e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1	jhu-clsp/mmBERT-base	0.082	m1	encoder	zephyrus	3	3	111M	20.5 ms	97.57% ±0.13	97.18% ±0.15	97.71% ±0.22	97.38% ±0.11	97.22% ±0.17	97.60% ±0.13	98.39% ±0.80	87.19% ±0.22	97.56% ±0.24	97.54% ±0.10	94.38% ±0.40	92.65% ±1.09	92.49% ±0.10	82.43% ±1.31	88.07% ±1.67	84.02% ±1.29	87.17% ±0.57	93.00% ±0.19	93.33% ±1.80	53.22% ±1.42	85.81% ±1.61	89.37% ±0.89	83.10% ±2.89	71.39% ±3.10	okp85_not_met	–
72	e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.067	m1	encoder	legion	1	1	105M	48.8 ms	97.42%	97.15%	97.49%	97.24%	97.17%	97.46%	98.39%	86.57%	97.41%	97.36%	89.96%	95.32%	91.18%	82.11%	89.48%	84.18%	86.21%	92.06%	92.86%	45.61%	86.13%	90.03%	82.14%	73.72%	okp85_not_met	–
73	e2_visobert_cross_ce_real_frozenemb_lr4e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	uitnlp/visobert	0.037	m1	encoder	zephyrus	1	1	86M	18.8 ms	97.54%	97.12%	97.53%	97.27%	97.35%	97.57%	98.39%	88.13%	97.27%	97.66%	94.38%	94.38%	89.43%	80.67%	86.84%	81.84%	83.81%	90.74%	93.57%	40.06%	84.62%	87.25%	87.86%	57.75%	okp85_not_met	–
74	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1	FacebookAI/xlm-roberta-base	0.027	m1	encoder	zephyrus	1	1	86M	7.8 ms	97.38%	97.17%	97.27%	97.14%	97.21%	97.41%	98.39%	87.19%	97.61%	97.75%	95.58%	93.70%	91.26%	81.79%	87.95%	83.75%	86.10%	91.98%	88.57%	47.69%	85.29%	88.55%	81.43%	66.28%	okp85_not_met	–
75	e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m1mmBERT base · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.057	m1	encoder	zephyrus	3	3	110M	36.2 ms	97.49% ±0.12	97.12% ±0.14	97.77% ±0.26	97.38% ±0.09	97.27% ±0.04	97.52% ±0.13	98.39% ±1.06	87.33% ±1.73	97.55% ±0.13	97.47% ±0.23	92.10% ±0.84	95.69% ±0.22	92.34% ±0.94	86.94% ±1.30	86.89% ±1.38	85.44% ±0.76	87.78% ±1.28	92.97% ±0.73	94.05% ±1.80	51.09% ±4.91	86.65% ±1.34	90.48% ±0.97	79.29% ±2.58	74.25% ±4.83	okp85_not_met	–
76	e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.206	m1	encoder	zephyrus	3	3	86M	18.8 ms	97.88% ±0.29	97.84% ±0.42	98.03% ±0.18	97.90% ±0.32	97.75% ±0.35	97.89% ±0.28	98.39% ±0.40	91.15% ±2.87	98.06% ±0.07	97.94% ±0.06	96.92% ±1.16	94.04% ±1.07	93.55% ±1.27	84.79% ±0.90	90.82% ±1.43	87.03% ±0.49	89.37% ±1.77	93.80% ±0.98	89.29% ±5.00	65.69% ±12.86	86.66% ±0.61	90.24% ±0.51	82.62% ±2.97	74.30% ±1.04	okp85_not_met	–
77	e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.034	m1	encoder	zephyrus	3	3	87M	20.6 ms	97.42% ±0.21	97.15% ±0.38	97.39% ±0.43	97.19% ±0.38	97.22% ±0.23	97.46% ±0.20	98.39% ±0.70	87.21% ±1.58	97.53% ±0.14	97.34% ±0.20	94.11% ±1.41	92.65% ±1.60	93.06% ±0.31	84.40% ±1.81	89.24% ±2.07	85.94% ±0.68	88.37% ±0.53	93.41% ±0.23	91.19% ±1.09	57.67% ±3.42	86.97% ±1.21	90.41% ±0.43	85.00% ±1.24	74.48% ±3.03	okp85_not_met	–
78	e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.052	m1	encoder	zephyrus	3	3	86M	19.4 ms	97.61% ±0.33	97.34% ±0.47	97.78% ±0.17	97.51% ±0.33	97.38% ±0.39	97.63% ±0.32	98.26% ±0.23	88.38% ±2.55	97.78% ±0.12	97.65% ±0.30	95.31% ±0.46	93.57% ±1.42	92.89% ±0.82	84.41% ±1.29	86.95% ±0.56	84.73% ±0.51	87.87% ±1.51	93.28% ±0.61	92.14% ±3.98	56.34% ±6.39	86.46% ±0.40	89.96% ±0.53	81.19% ±1.09	75.28% ±0.19	oktest_informedp85_not_met	–
79	e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.033	m1	encoder	zephyrus	3	3	135M	20.4 ms	97.42% ±0.37	97.18% ±0.38	97.35% ±0.30	97.19% ±0.34	97.16% ±0.37	97.46% ±0.36	98.26% ±0.23	87.00% ±2.02	97.34% ±0.14	97.17% ±0.24	93.17% ±0.40	93.18% ±1.14	92.73% ±0.64	84.76% ±1.48	86.81% ±1.12	84.77% ±0.72	87.58% ±1.30	93.13% ±0.50	93.10% ±2.30	55.68% ±4.29	86.91% ±0.58	89.87% ±0.73	81.67% ±1.09	76.06% ±1.07	oktest_informedp85_not_met	–
80	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1	jhu-clsp/mmBERT-base	0.144	m1	encoder	zephyrus	3	3	111M	18.3 ms	97.66% ±0.27	97.43% ±0.35	97.59% ±0.24	97.45% ±0.22	97.37% ±0.28	97.69% ±0.27	98.26% ±1.41	88.49% ±2.39	97.72% ±0.05	97.69% ±0.08	94.91% ±1.67	93.92% ±0.75	92.16% ±0.51	82.40% ±0.38	87.88% ±1.52	84.01% ±0.77	86.99% ±0.59	92.76% ±0.38	92.14% ±0.71	51.64% ±3.49	86.14% ±1.78	89.45% ±1.44	83.57% ±4.34	69.71% ±8.49	okp85_not_met	–
81	e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.082	m1	encoder	zephyrus	3	3	111M	18.6 ms	97.89% ±0.49	97.56% ±0.68	98.03% ±0.44	97.75% ±0.57	97.66% ±0.61	97.92% ±0.48	98.26% ±1.29	88.91% ±3.07	97.71% ±0.73	97.76% ±0.63	93.57% ±3.87	93.81% ±0.96	92.21% ±0.96	81.95% ±2.00	91.96% ±0.52	85.21% ±1.27	87.04% ±1.24	92.82% ±0.69	91.67% ±4.31	52.31% ±7.01	85.35% ±0.41	89.45% ±0.49	80.95% ±2.97	72.33% ±2.89	cần sửap85_not_met	–
82	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1	jhu-clsp/mmBERT-base	0.108	m1	encoder	zephyrus	3	3	111M	18.7 ms	97.70% ±0.20	97.39% ±0.34	97.56% ±0.25	97.41% ±0.31	97.38% ±0.38	97.73% ±0.19	98.26% ±0.61	87.74% ±2.32	97.63% ±0.17	97.68% ±0.26	93.57% ±2.81	93.61% ±1.46	93.15% ±0.50	83.78% ±1.65	89.12% ±0.39	85.56% ±1.15	88.17% ±0.67	93.51% ±0.33	92.38% ±2.18	58.10% ±5.14	86.92% ±1.11	90.27% ±0.29	85.24% ±3.22	77.23% ±1.87	okp85_not_met	–
83	e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-0.6B-Base	0.075	m1	decoder	legion	3	3	10M	45.9 ms	97.53% ±0.29	96.93% ±0.41	97.92% ±0.23	97.37% ±0.27	97.26% ±0.41	97.56% ±0.28	98.13% ±1.29	87.71% ±2.13	97.70% ±0.09	97.65% ±0.15	93.57% ±1.75	94.85% ±0.85	92.14% ±1.17	82.34% ±1.04	91.13% ±1.76	85.22% ±1.49	87.68% ±1.92	92.81% ±0.87	94.05% ±1.49	50.70% ±6.26	86.81% ±0.82	90.97% ±0.78	82.62% ±3.30	76.00% ±1.85	okp85_not_met	–
84	e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1	Qwen/Qwen3-0.6B-Base	0.077	m1	encoder	legion	3	3	97M	50.7 ms	97.10% ±0.18	96.99% ±0.33	97.04% ±0.13	96.94% ±0.20	96.80% ±0.24	97.14% ±0.17	97.99% ±1.61	86.78% ±2.11	97.11% ±0.09	96.95% ±0.25	90.90% ±1.62	94.83% ±0.62	90.63% ±0.96	83.36% ±1.96	88.20% ±1.60	84.39% ±2.17	86.14% ±0.93	91.50% ±0.68	90.95% ±3.60	45.69% ±4.25	86.63% ±1.72	89.43% ±0.08	76.90% ±1.80	71.78% ±1.55	cần sửap85_not_met	–
85	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_m1	vinai/phobert-base-v2	0.093	m1	encoder	zephyrus	3	3	86M	20.4 ms	97.33% ±0.25	96.95% ±0.36	97.68% ±0.16	97.25% ±0.23	97.02% ±0.28	97.36% ±0.24	97.99% ±0.40	87.05% ±1.23	97.74% ±0.33	97.65% ±0.26	94.91% ±1.23	95.05% ±0.93	92.95% ±0.20	83.49% ±2.36	91.63% ±2.20	86.32% ±0.46	88.13% ±0.23	93.39% ±0.20	93.10% ±0.41	55.88% ±1.25	87.31% ±1.19	90.40% ±0.71	85.00% ±3.78	75.16% ±4.91	okp85_not_met	–
86	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m1	vinai/phobert-base-v2	0.104	m1	encoder	zephyrus	3	3	86M	22.0 ms	97.66% ±0.28	97.38% ±0.41	97.65% ±0.34	97.46% ±0.38	97.38% ±0.37	97.69% ±0.27	97.99% ±1.45	88.27% ±2.74	97.54% ±0.15	97.42% ±0.03	94.11% ±0.23	93.24% ±0.37	92.98% ±1.12	85.84% ±1.52	88.95% ±2.31	86.38% ±1.53	88.05% ±2.09	93.34% ±0.84	92.14% ±2.86	58.77% ±8.67	88.46% ±0.79	90.50% ±0.41	85.24% ±1.09	76.92% ±3.19	okp85_not_met	–
87	e2_xlmr_base_cross_ce_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.044	m1	encoder	zephyrus	3	3	86M	12.1 ms	97.38% ±0.26	97.06% ±0.21	97.11% ±0.30	97.00% ±0.29	97.21% ±0.29	97.42% ±0.25	97.99% ±0.40	87.08% ±2.15	96.84% ±0.23	97.20% ±0.29	92.37% ±1.06	92.39% ±1.25	91.55% ±1.24	81.94% ±0.88	87.22% ±2.07	83.37% ±1.94	86.23% ±1.58	92.19% ±0.95	90.95% ±3.30	50.55% ±7.61	84.59% ±1.94	88.52% ±1.21	81.90% ±5.46	68.09% ±10.26	okp85_not_met	–
88	e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.019	m1	encoder	zephyrus	3	3	86M	10.9 ms	97.49% ±0.15	97.03% ±0.06	97.85% ±0.25	97.38% ±0.17	97.28% ±0.22	97.52% ±0.14	97.99% ±0.70	87.70% ±1.98	97.36% ±0.28	97.54% ±0.10	93.98% ±1.06	93.49% ±0.88	91.76% ±0.80	81.82% ±1.03	89.87% ±1.12	84.56% ±1.07	86.67% ±1.03	92.33% ±0.54	93.10% ±2.97	52.15% ±5.51	84.81% ±0.31	89.13% ±0.48	82.14% ±3.57	71.21% ±5.94	okp85_not_met	–
89	e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.049	m1	encoder	zephyrus	3	3	86M	18.9 ms	97.23% ±0.41	96.89% ±0.43	97.21% ±0.48	96.96% ±0.41	96.89% ±0.48	97.27% ±0.41	97.99% ±1.75	85.91% ±0.51	97.42% ±0.70	97.37% ±0.53	94.38% ±2.13	93.03% ±1.65	92.24% ±0.31	81.78% ±1.27	89.06% ±1.38	84.15% ±0.65	87.32% ±0.50	92.79% ±0.34	93.10% ±2.06	52.29% ±0.98	85.23% ±0.45	89.92% ±0.32	84.29% ±2.58	71.54% ±0.94	okp85_not_met	–
90	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.159	m1	encoder	zephyrus	3	3	86M	22.1 ms	97.09% ±0.20	96.64% ±0.57	96.87% ±0.27	96.68% ±0.41	96.87% ±0.32	97.12% ±0.20	97.72% ±1.16	86.91% ±0.81	96.78% ±0.48	96.97% ±0.34	93.31% ±0.61	92.00% ±2.36	92.56% ±1.04	83.02% ±2.74	86.06% ±5.16	83.26% ±4.82	87.78% ±1.38	92.93% ±0.80	91.67% ±1.49	56.95% ±8.27	86.76% ±0.96	89.49% ±0.29	85.71% ±3.78	70.10% ±1.53	okp85_not_met	–
91	e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-4B	0.061	m1	decoder	kaggle	2	2	79M	126.5 ms	96.23% ±1.13	96.17% ±0.49	96.75% ±0.80	96.38% ±0.67	95.98% ±0.79	96.28% ±1.13	97.39% ±0.85	85.39% ±0.10	96.45% ±0.88	96.14% ±1.16	90.76% ±4.54	93.48% ±2.43	90.59% ±0.22	82.54% ±4.30	86.37% ±2.39	82.51% ±1.73	84.71% ±0.86	91.23% ±0.33	90.71% ±3.03	48.84% ±1.10	85.61% ±2.06	87.42% ±1.78	71.79% ±12.63	75.53% ±6.32	okp85_not_met	–
92	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.025	m1	encoder	zephyrus	1	1	86M	11.5 ms	97.31%	97.01%	96.37%	96.59%	97.06%	97.34%	97.19%	86.74%	97.02%	97.06%	92.37%	93.12%	92.46%	79.41%	77.99%	76.53%	86.83%	92.74%	90.00%	55.02%	82.15%	88.57%	80.71%	71.07%	okp85_not_met	–
93	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.152	m1	decoder	zephyrus	3	3	13M	62.5 ms	96.88% ±0.23	96.66% ±0.26	97.34% ±0.21	96.95% ±0.20	96.76% ±0.24	96.90% ±0.22	97.05% ±1.23	88.02% ±1.88	96.99% ±0.27	97.01% ±0.38	93.04% ±1.41	93.17% ±1.73	91.27% ±0.80	80.36% ±0.75	87.35% ±0.60	82.67% ±0.43	85.25% ±1.33	91.72% ±0.70	85.24% ±0.41	53.08% ±4.73	84.16% ±1.01	86.78% ±0.43	75.48% ±2.89	67.73% ±1.30	okp85_not_met	–
94	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m1	vinai/phobert-base-v2	0.035	m1	encoder	zephyrus	3	3	86M	20.7 ms	97.35% ±0.10	96.87% ±0.33	97.15% ±0.30	96.95% ±0.32	96.92% ±0.33	97.38% ±0.09	96.92% ±1.85	86.72% ±1.78	97.61% ±0.60	97.50% ±0.55	93.31% ±3.73	94.94% ±1.39	92.51% ±0.99	83.35% ±0.74	89.61% ±0.83	85.37% ±0.78	87.68% ±1.10	93.04% ±0.76	94.05% ±1.49	53.77% ±5.56	87.18% ±0.29	90.40% ±0.91	83.33% ±3.93	74.18% ±4.04	cần sửap85_not_met	–
95	e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-base	0.052	m1	encoder	zephyrus	1	1	86M	11.1 ms	96.72%	96.23%	96.72%	96.41%	96.49%	96.75%	96.79%	86.07%	95.57%	96.22%	89.56%	91.39%	91.29%	78.74%	79.11%	77.03%	85.77%	91.90%	90.71%	48.85%	81.25%	86.47%	76.43%	59.12%	okp85_not_met	–
96	e2_visobert_cross_ce_real_frozenemb_lr2e5_m1	uitnlp/visobert	0.027	m1	encoder	zephyrus	1	1	86M	17.1 ms	96.99%	96.75%	96.80%	96.73%	96.88%	97.02%	96.39%	87.59%	97.06%	97.26%	91.16%	96.19%	89.62%	79.10%	84.00%	80.01%	84.35%	90.64%	90.71%	42.47%	83.29%	87.64%	79.29%	64.91%	okp85_not_met	–
97	e2_visobert_cross_ce_real_frozenemb_lr1e5_m1	uitnlp/visobert	0.186	m1	encoder	zephyrus	1	1	86M	12.5 ms	96.10%	96.29%	96.14%	96.16%	96.04%	96.12%	95.18%	86.81%	96.28%	96.18%	91.57%	91.20%	89.62%	78.41%	83.53%	79.66%	83.28%	90.45%	87.14%	44.69%	80.69%	84.52%	83.57%	51.32%	okp85_not_met	–
98	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1	FacebookAI/xlm-roberta-base	0.161	m1	encoder	zephyrus	1	1	86M	12.5 ms	96.68%	96.50%	96.10%	96.23%	96.45%	96.71%	95.18%	86.81%	96.27%	96.51%	92.37%	90.55%	91.26%	78.09%	84.34%	80.15%	85.81%	91.75%	90.00%	52.72%	80.90%	86.71%	87.14%	58.10%	okp85_not_met	–
99	e1_tfidf_hybrid_qwen3emb_lr_m1TF-IDF câu KH (char 1-5, word 1-2-gram, giữ từ 1 ký tự, thêm kênh char bỏ dấu) + câu bot (char 2-4) + tương tác loại câu bot × từ câu KH · vector đóng băng Qwen3-Embedding-0.6B (cặp bot + KH, pooling last_token), trọng số dense 1 · LogReg · train thật, nhãn m1	Qwen/Qwen3-Embedding-0.6B	0.149	m1	tfidf_hybrid	zephyrus	1	1	157,386	35.1 ms	96.06%	96.28%	96.62%	96.41%	95.85%	96.09%	94.38%	85.77%	96.32%	95.59%	86.75%	92.31%	90.09%	82.15%	82.69%	81.30%	83.65%	90.69%	85.71%	47.06%	84.25%	86.29%	75.00%	64.02%	okp85_not_met	–
100	e1_tfidf_svmcal_ce_synth1x_fbmatched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	TF-IDF	0.207	m1	tfidf	zephyrus	1	1	117,072	4.3 ms	95.82%	96.22%	95.82%	95.99%	95.67%	95.85%	92.77%	86.84%	95.86%	95.52%	86.35%	92.27%	91.68%	84.99%	84.00%	83.87%	86.10%	91.93%	82.86%	57.43%	85.98%	87.18%	70.71%	73.33%	oktest_informedp85_not_met	–
101	e1_tfidf_svmcal_ce_synth1x_nofb_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent omi_intent_repeat, callback, relationship, negate, affirm), nhãn m1	TF-IDF	0.164	m1	tfidf	zephyrus	1	1	116,442	4.2 ms	95.75%	96.03%	95.86%	95.91%	95.57%	95.77%	92.77%	85.87%	95.94%	95.61%	85.54%	93.42%	91.53%	83.45%	83.81%	83.03%	85.72%	91.84%	82.14%	55.02%	85.99%	87.19%	69.29%	74.62%	oktest_informedp85_not_met	–
102	e1_tfidf_svmcal_ce_synth1x_noqparticle_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm q_particle), nhãn m1	TF-IDF	0.280	m1	tfidf	zephyrus	1	1	127,464	4.5 ms	95.63%	96.06%	95.66%	95.83%	95.47%	95.66%	92.37%	85.82%	96.22%	95.94%	89.16%	92.12%	90.32%	82.98%	86.46%	83.58%	84.29%	90.98%	85.71%	47.24%	84.32%	86.36%	79.29%	58.12%	oktest_informedp85_not_met	–
103	e1_tfidf_svmcal_ce_real_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn m1	TF-IDF	0.178	m1	tfidf	zephyrus	1	1	111,636	4.5 ms	95.78%	96.15%	95.75%	95.92%	95.59%	95.81%	91.97%	86.42%	95.87%	95.53%	85.14%	92.58%	91.53%	83.33%	84.34%	83.14%	85.86%	91.89%	85.00%	54.84%	85.80%	87.58%	72.86%	74.45%	okp85_not_met	–
104	e1_tfidf_svmcal_ce_synth1x_nofiller_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm filler_only), nhãn m1	TF-IDF	0.267	m1	tfidf	zephyrus	1	1	128,538	4.2 ms	95.67%	95.99%	95.70%	95.81%	95.45%	95.70%	91.97%	85.13%	96.17%	95.88%	89.16%	91.36%	90.40%	83.94%	85.27%	83.32%	84.33%	91.05%	88.57%	47.51%	85.05%	86.67%	82.86%	59.49%	oktest_informedp85_not_met	–
105	e1_tfidf_svmcal_ce_synth1x_noofftopic_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm prefix_offtopic), nhãn m1	TF-IDF	0.298	m1	tfidf	zephyrus	1	1	127,548	4.3 ms	95.63%	96.05%	95.66%	95.82%	95.47%	95.66%	91.97%	85.45%	96.21%	95.93%	88.35%	92.05%	90.48%	82.10%	86.69%	83.31%	84.58%	91.07%	85.71%	48.58%	84.56%	86.65%	81.43%	59.38%	oktest_informedp85_not_met	–
106	e2_dual_qwen3emb06b_concat_ce_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.227	m1	decoder	zephyrus	3	3	2M	54.5 ms	95.56% ±0.08	95.62% ±0.15	95.61% ±0.14	95.59% ±0.15	95.13% ±0.09	95.59% ±0.08	91.83% ±0.23	86.84% ±0.93	95.75% ±0.24	95.38% ±0.29	87.28% ±1.62	93.70% ±1.06	88.87% ±0.69	84.30% ±1.88	81.12% ±2.46	80.98% ±1.06	81.73% ±0.98	89.46% ±0.55	86.90% ±1.09	46.30% ±3.09	83.03% ±1.29	83.77% ±1.01	80.95% ±0.41	59.44% ±2.97	okp85_not_met	–
107	e1_tfidf_svmcal_ce_synth1x_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1, nhãn m1	TF-IDF	0.282	m1	tfidf	zephyrus	1	1	128,604	4.4 ms	95.67%	96.01%	95.59%	95.77%	95.40%	95.70%	91.57%	85.39%	96.12%	95.82%	88.76%	91.70%	90.21%	81.82%	86.26%	82.88%	84.07%	90.86%	85.00%	47.04%	85.00%	86.62%	82.86%	59.18%	oktest_informedp85_not_met	–
108	e1_tfidf_svmcal_ce_synth1x_matched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ intent của train thật), nhãn m1	TF-IDF	0.190	m1	tfidf	zephyrus	1	1	112,440	4.3 ms	95.63%	95.99%	95.58%	95.75%	95.38%	95.65%	91.57%	86.04%	95.90%	95.39%	85.14%	92.17%	91.49%	83.32%	84.05%	83.01%	85.70%	91.83%	84.29%	55.14%	86.21%	87.46%	71.43%	74.07%	oktest_informedp85_not_met	–
109	e1_tfidf_svmcal_ce_synth1x_nomoney_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm money), nhãn m1	TF-IDF	0.310	m1	tfidf	zephyrus	1	1	127,230	4.1 ms	95.71%	96.17%	95.62%	95.87%	95.53%	95.73%	91.57%	86.69%	96.03%	95.72%	87.55%	92.37%	90.63%	83.15%	87.03%	84.09%	84.90%	91.20%	86.43%	49.39%	84.02%	86.01%	80.00%	56.28%	oktest_informedp85_not_met	–
110	e1_tfidf_svmcal_ce_synth1x_fbonly_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent nlu_fallback), nhãn m1	TF-IDF	0.356	m1	tfidf	zephyrus	1	1	121,248	4.2 ms	95.90%	96.43%	95.73%	96.06%	95.75%	95.91%	91.16%	87.98%	96.07%	95.77%	89.56%	89.56%	89.74%	81.46%	86.63%	82.70%	83.86%	90.57%	86.43%	44.65%	83.68%	85.04%	84.29%	50.43%	oktest_informedp85_not_met	–
–	e2_bamibert_cross_ce_real_frozenemb_lr1e5_m1	–	–	m1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e2_bamibert_cross_ce_real_frozenemb_lr2e5_m1	–	–	m1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e2_bamibert_cross_ce_real_frozenemb_lr4e5_m1	–	–	m1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e3_fbverifier_devtop1_familygeomean_fbf1_m1	–	hard	m1	ensemble	zephyrus	1	1	N/A	271.8 ms	98.48%	98.12%	98.32%	98.22%	98.33%	98.48%	N/A	94.84%	98.22%	98.33%	95.98%	94.84%	95.69%	87.21%	92.80%	89.56%	92.30%	95.71%	N/A	81.21%	89.56%	92.30%	86.43%	81.21%	ok	–
–	e3_fbverifier_devtop1_familygeomean_macrof1_m1	–	hard	m1	ensemble	zephyrus	1	1	N/A	271.8 ms	98.48%	98.12%	98.32%	98.22%	98.33%	98.48%	N/A	94.84%	98.22%	98.33%	95.98%	94.84%	95.69%	87.21%	92.80%	89.56%	92.30%	95.71%	N/A	81.21%	89.56%	92.30%	86.43%	81.21%	ok	–
1	e3_ensemble_mean_all	–	0.065	v1	ensemble	zephyrus	1	1	N/A	429.8 ms	97.80%	97.15%	97.44%	97.20%	97.45%	97.83%	98.85%	87.50%	98.11%	98.04%	95.04%	95.77%	92.79%	82.95%	88.69%	84.83%	87.51%	93.21%	90.20%	55.87%	86.72%	89.87%	83.01%	73.41%	okp85_not_met	–
2	e3_ensemble_greedy_all	–	0.070	v1	ensemble	zephyrus	1	1	N/A	429.8 ms	97.69%	96.77%	97.72%	97.18%	97.24%	97.73%	98.47%	86.87%	98.16%	98.10%	95.04%	95.77%	92.67%	82.73%	88.39%	84.57%	87.20%	93.10%	89.54%	55.47%	86.83%	90.00%	82.35%	74.56%	okp85_not_met	–
3	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5	–	0.023	v1	encoder	zephyrus	1	1	86M	11.1 ms	97.77%	97.23%	97.44%	97.29%	97.54%	97.79%	98.09%	89.55%	96.90%	97.04%	91.22%	94.47%	92.94%	80.39%	87.57%	82.72%	87.14%	93.23%	86.27%	60.55%	83.59%	88.09%	81.05%	69.27%	okp85_not_met	–
4	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1	vinai/phobert-base-v2	0.011	v1	encoder	zephyrus	1	1	86M	11.9 ms	97.62%	97.10%	97.23%	97.11%	97.17%	97.65%	97.71%	88.58%	98.05%	97.97%	95.42%	97.66%	92.44%	82.30%	85.16%	83.06%	87.37%	92.76%	86.27%	56.65%	85.56%	88.93%	80.39%	70.69%	okp85_not_met	–
5	e2_xlmr_base_cross_wcesqrt_real_frozenemb	FacebookAI/xlm-roberta-base	0.021	v1	encoder	zephyrus	3	3	86M	16.1 ms	97.19% ±0.16	96.61% ±0.10	97.17% ±0.09	96.80% ±0.10	96.84% ±0.12	97.23% ±0.15	97.71% ±0.38	85.62% ±0.52	97.23% ±0.12	97.41% ±0.18	92.88% ±1.45	94.33% ±1.35	92.50% ±0.34	82.06% ±1.22	87.96% ±0.44	84.16% ±0.89	87.17% ±0.77	92.88% ±0.28	87.58% ±1.31	56.71% ±2.55	84.77% ±1.14	88.54% ±0.82	81.05% ±1.31	69.50% ±2.97	okp85_not_met	–
6	e3_prior_em_mmbert_bottype	–	0.082	v1	post_hoc_transductive	zephyrus	1	1	N/A	24.8 ms	97.55%	96.93%	97.26%	97.02%	97.05%	97.59%	97.71%	86.49%	97.93%	97.83%	93.51%	95.33%	87.17%	74.44%	79.79%	74.17%	75.67%	88.28%	85.62%	62.38%	73.61%	76.03%	75.16%	76.16%	oktransductivep85_not_met	–
7	e3_prior_em_mmbert_global	–	0.082	v1	post_hoc_transductive	zephyrus	1	1	N/A	24.8 ms	97.55%	96.93%	97.26%	97.02%	97.05%	97.59%	97.71%	86.49%	97.93%	97.83%	93.51%	95.33%	92.48%	85.70%	86.72%	85.46%	86.56%	92.74%	88.24%	57.69%	87.09%	88.51%	81.05%	75.61%	oktransductivep85_not_met	–
8	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5	jhu-clsp/mmBERT-base	0.061	v1	encoder	zephyrus	3	3	111M	24.8 ms	97.54% ±0.18	96.98% ±0.23	97.08% ±0.24	96.95% ±0.18	97.09% ±0.22	97.57% ±0.18	97.58% ±0.58	87.26% ±0.40	97.56% ±0.13	97.50% ±0.09	92.24% ±2.10	95.05% ±1.43	92.72% ±0.52	83.66% ±0.26	90.04% ±0.84	85.79% ±0.29	87.14% ±0.86	93.12% ±0.44	88.24% ±0.65	55.89% ±3.08	87.12% ±0.79	89.05% ±0.67	81.05% ±2.61	70.95% ±2.31	cần sửap85_not_met	–
9	e2_xlmr_base_cross_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	encoder	zephyrus	3	3	86M	10.8 ms	97.47% ±0.48	96.79% ±0.43	97.06% ±0.21	96.88% ±0.32	97.25% ±0.50	97.48% ±0.47	97.46% ±0.22	89.35% ±2.96	97.01% ±0.33	97.35% ±0.40	92.75% ±2.02	95.29% ±0.45	92.93% ±0.15	83.08% ±1.52	88.16% ±0.41	84.95% ±0.85	87.60% ±0.05	93.18% ±0.13	86.27% ±2.61	60.95% ±0.90	84.74% ±0.68	88.34% ±0.44	79.08% ±0.65	70.68% ±2.10	cần sửap85_not_met	high
10	e2_xlmr_base_cross_logitadjust_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	encoder	zephyrus	3	3	86M	14.7 ms	97.44% ±0.12	96.55% ±0.04	97.44% ±0.35	96.92% ±0.19	97.09% ±0.15	97.47% ±0.12	97.46% ±0.58	86.57% ±1.19	97.37% ±0.29	97.71% ±0.12	94.66% ±0.00	93.82% ±0.41	92.55% ±0.35	81.29% ±0.71	88.95% ±1.26	83.77% ±0.78	87.47% ±0.80	92.98% ±0.35	87.80% ±2.30	55.66% ±1.10	84.18% ±0.68	89.16% ±0.51	82.14% ±1.36	68.21% ±1.58	okp85_not_met	–
11	e2_qwen3_06b_base_cross_ce_real_lora	Qwen/Qwen3-0.6B-Base	0.098	v1	decoder	legion	1	1	10M	52.7 ms	97.30%	97.04%	97.24%	97.10%	96.99%	97.32%	97.33%	89.16%	97.33%	97.10%	91.22%	96.37%	92.52%	82.85%	88.19%	84.81%	87.49%	92.81%	87.58%	60.09%	84.66%	88.26%	79.08%	72.02%	okp85_not_met	–
12	e3_ensemble_stack_all	–	0.088	v1	ensemble	zephyrus	1	1	N/A	429.8 ms	97.26%	96.39%	96.52%	96.36%	96.75%	97.30%	97.33%	85.00%	96.86%	97.02%	91.22%	93.73%	92.63%	84.14%	87.87%	85.22%	87.45%	93.00%	88.24%	57.45%	85.07%	89.14%	81.70%	72.25%	okp85_not_met	–
13	e2_qwen3_17b_base_cross_ce_real_lora	Qwen/Qwen3-1.7B-Base	0.083	v1	decoder	legion	3	3	17M	53.1 ms	97.09% ±0.14	96.41% ±0.16	96.67% ±0.60	96.46% ±0.38	96.59% ±0.23	97.13% ±0.14	97.20% ±0.58	85.65% ±0.98	97.04% ±0.56	97.13% ±0.44	92.11% ±1.45	95.33% ±2.54	92.50% ±0.19	83.51% ±0.31	89.84% ±1.69	85.56% ±0.66	87.25% ±0.25	92.90% ±0.12	86.93% ±1.31	55.59% ±2.51	86.66% ±0.84	89.40% ±0.49	80.83% ±1.64	71.98% ±2.86	okp85_not_met	–
14	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5	vinai/phobert-base-v2	0.022	v1	encoder	zephyrus	3	3	86M	12.6 ms	97.33% ±0.16	97.00% ±0.31	96.56% ±0.54	96.71% ±0.40	97.14% ±0.22	97.34% ±0.15	97.07% ±0.96	89.99% ±1.66	97.36% ±0.29	97.30% ±0.19	92.88% ±0.44	96.56% ±0.23	92.98% ±0.10	84.30% ±0.77	86.01% ±1.49	84.64% ±1.16	87.72% ±0.34	93.25% ±0.10	87.15% ±1.00	59.28% ±0.95	85.31% ±0.27	89.05% ±0.65	79.52% ±0.38	71.05% ±2.00	cần sửap85_not_met	–
15	e2_me5_base_cross_ce_real_frozenemb_lr1e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	intfloat/multilingual-e5-base	0.088	v1	encoder	zephyrus	1	1	86M	11.3 ms	97.12%	96.52%	96.85%	96.63%	96.91%	97.13%	96.95%	88.50%	96.17%	96.79%	90.46%	94.42%	92.29%	80.81%	86.69%	82.85%	87.01%	92.63%	89.54%	58.80%	84.07%	87.94%	81.05%	70.06%	okp85_not_met	–
16	e2_me5_base_cross_ce_real_frozenemb_lr4e5	–	0.028	v1	encoder	zephyrus	1	1	86M	11.6 ms	97.23%	96.42%	96.82%	96.54%	96.80%	97.26%	96.95%	85.81%	97.17%	96.91%	91.60%	92.66%	92.90%	84.46%	88.31%	85.81%	87.59%	93.17%	88.24%	60.00%	85.57%	88.40%	83.01%	70.17%	okp85_not_met	–
17	e2_xlmr_base_khonly_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.076	v1	encoder	zephyrus	3	3	86M	8.5 ms	96.89% ±0.33	96.68% ±0.32	96.75% ±0.43	96.67% ±0.38	96.74% ±0.37	96.90% ±0.33	96.69% ±0.58	89.11% ±1.40	96.72% ±0.22	96.74% ±0.21	91.98% ±1.53	94.26% ±0.86	92.71% ±0.51	83.88% ±0.46	87.66% ±1.03	85.21% ±0.71	87.60% ±0.45	92.97% ±0.44	86.06% ±2.30	60.29% ±3.36	85.46% ±0.20	88.93% ±0.52	81.48% ±1.51	71.39% ±2.23	okp85_not_met	–
18	e2_me5_base_cross_ce_real_frozenemb_lr2e5	intfloat/multilingual-e5-base	0.056	v1	encoder	zephyrus	3	3	86M	10.9 ms	97.33% ±0.16	96.86% ±0.34	97.17% ±0.46	96.96% ±0.40	97.04% ±0.08	97.35% ±0.16	96.56% ±0.76	88.57% ±0.82	97.24% ±0.12	97.20% ±0.14	92.49% ±0.96	94.32% ±1.87	92.89% ±0.28	82.93% ±1.72	87.59% ±0.94	84.67% ±1.38	87.57% ±0.54	93.20% ±0.24	87.36% ±1.36	59.69% ±1.44	84.14% ±1.04	88.05% ±0.45	79.74% ±2.85	68.74% ±1.82	cần sửap85_not_met	–
19	e2_gemma3_270m_cross_ce_real_lora	unsloth/gemma-3-270m	0.116	v1	decoder	zephyrus	3	3	4M	54.2 ms	97.02% ±0.18	96.64% ±0.37	96.96% ±0.24	96.75% ±0.30	96.63% ±0.36	97.05% ±0.17	96.44% ±0.88	87.65% ±2.07	96.78% ±0.48	96.76% ±0.43	91.22% ±2.38	93.60% ±0.58	92.26% ±0.39	81.57% ±0.44	87.85% ±0.82	83.56% ±0.65	86.48% ±0.27	92.65% ±0.27	88.24% ±2.36	56.64% ±2.95	84.05% ±0.55	87.78% ±0.30	79.30% ±1.00	69.60% ±0.99	okp85_not_met	–
20	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen	Qwen/Qwen3-Embedding-0.6B	0.114	v1	decoder	zephyrus	3	3	13M	82.5 ms	96.65% ±0.13	96.42% ±0.10	96.64% ±0.18	96.47% ±0.14	96.35% ±0.21	96.68% ±0.13	96.06% ±1.23	86.59% ±0.99	96.60% ±0.16	96.50% ±0.18	91.22% ±1.01	92.64% ±0.65	91.99% ±0.14	81.95% ±0.67	88.80% ±0.05	84.31% ±0.60	86.47% ±0.17	92.35% ±0.10	84.97% ±2.85	56.06% ±0.80	85.91% ±0.36	87.57% ±0.44	76.91% ±1.89	69.08% ±1.10	okp85_not_met	–
21	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5	–	0.075	v1	encoder	zephyrus	1	1	111M	51.6 ms	97.51%	97.12%	96.77%	96.88%	97.06%	97.54%	95.80%	87.46%	97.18%	97.25%	91.98%	92.69%	92.94%	83.33%	88.12%	84.85%	87.54%	93.31%	87.58%	56.78%	85.14%	89.27%	82.35%	69.61%	okp85_not_met	–
22	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5	–	0.026	v1	encoder	zephyrus	1	1	86M	12.9 ms	96.65%	96.26%	94.71%	95.29%	96.38%	96.66%	95.80%	87.15%	96.67%	96.78%	91.60%	95.24%	92.07%	88.35%	76.76%	76.93%	86.60%	92.28%	88.89%	56.43%	84.44%	87.99%	79.74%	70.93%	okp85_not_met	–
23	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	FacebookAI/xlm-roberta-base	0.123	v1	encoder	zephyrus	1	1	86M	10.2 ms	96.54%	95.48%	94.64%	94.94%	96.40%	96.55%	95.80%	86.55%	95.06%	96.18%	90.84%	90.15%	92.56%	77.16%	76.38%	76.08%	87.09%	92.73%	84.31%	60.00%	81.67%	87.64%	79.08%	66.85%	okp85_not_met	–
24	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5	jhu-clsp/mmBERT-base	0.120	v1	encoder	zephyrus	2	2	111M	23.9 ms	97.15% ±0.10	96.18% ±0.14	96.98% ±0.20	96.53% ±0.16	96.53% ±0.09	97.19% ±0.10	95.61% ±0.27	86.08% ±0.18	97.28% ±0.41	97.28% ±0.39	92.75% ±1.08	94.37% ±1.36	92.81% ±0.13	81.63% ±1.25	88.31% ±0.61	84.03% ±0.51	87.69% ±0.31	93.10% ±0.14	85.29% ±1.39	60.29% ±0.79	84.27% ±0.28	88.39% ±0.57	77.12% ±0.92	69.90% ±3.76	cần sửap85_not_met	–
25	e2_visobert_cross_ce_real_frozenemb_lr2e5	–	0.093	v1	encoder	zephyrus	1	1	86M	11.8 ms	96.04%	95.72%	95.89%	95.75%	95.85%	96.06%	95.04%	85.86%	95.99%	95.97%	89.69%	92.52%	92.22%	80.91%	86.94%	83.04%	86.31%	92.54%	83.66%	58.18%	84.05%	87.53%	81.05%	68.13%	okp85_not_met	–
26	e2_visobert_cross_ce_real_frozenemb_lr4e5	–	0.066	v1	encoder	zephyrus	1	1	86M	16.3 ms	96.58%	96.18%	96.08%	96.09%	96.26%	96.59%	94.66%	87.94%	96.14%	96.45%	90.46%	93.31%	92.07%	82.45%	88.73%	84.64%	86.40%	92.46%	84.97%	55.56%	85.58%	88.03%	81.05%	67.39%	okp85_not_met	–
27	e1_tfidf_svmcal_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn v1	TF-IDF	0.123	v1	tfidf	zephyrus	1	1	119,328	5.4 ms	95.42%	95.44%	95.80%	95.57%	95.31%	95.46%	94.27%	85.17%	95.87%	95.67%	87.79%	95.04%	91.88%	84.06%	86.93%	84.58%	86.11%	92.29%	88.89%	54.40%	86.51%	88.43%	74.51%	74.03%	okp85_not_met	–
28	e1_tfidf_logreg_balanced_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg cân bằng lớp · train thật, nhãn v1	TF-IDF	0.256	v1	tfidf	zephyrus	1	1	119,328	3.6 ms	95.50%	95.50%	96.21%	95.82%	95.45%	95.52%	93.89%	86.62%	96.17%	95.72%	90.84%	91.89%	90.39%	78.70%	87.99%	81.67%	84.25%	90.92%	86.27%	50.00%	82.92%	85.76%	79.74%	59.51%	okp85_not_met	–
29	e1_tfidf_logreg_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật, nhãn v1	TF-IDF	0.125	v1	tfidf	zephyrus	1	1	119,328	3.6 ms	95.53%	95.52%	95.69%	95.56%	95.46%	95.56%	93.89%	86.01%	95.83%	95.62%	87.02%	94.61%	90.93%	81.78%	85.92%	82.73%	84.46%	91.45%	88.24%	50.94%	84.60%	86.71%	74.51%	67.46%	okp85_not_met	–
30	e2_visobert_cross_ce_real_frozenemb_lr1e5	–	0.232	v1	encoder	zephyrus	1	1	86M	11.1 ms	95.17%	94.57%	94.96%	94.74%	95.10%	95.18%	92.75%	88.04%	94.77%	95.14%	90.46%	90.80%	91.80%	79.12%	85.57%	81.37%	85.15%	92.06%	78.43%	58.54%	81.19%	85.30%	73.86%	62.78%	okp85_not_met	–
31	e1_tfidf_logreg_ce_synth1x_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật + synth ×1, nhãn v1	TF-IDF	0.234	v1	tfidf	zephyrus	1	1	135,678	3.6 ms	95.21%	95.38%	94.99%	95.14%	95.12%	95.23%	92.37%	85.82%	95.79%	95.74%	90.08%	92.91%	91.34%	81.77%	85.78%	82.85%	85.14%	91.70%	84.97%	54.62%	83.94%	86.44%	75.82%	64.09%	oktest_informedp85_not_met	–
–	e2_mdeberta_cross_ce_real_frozenemb_lr1e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	–	–	v1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5	–	–	v1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e2_mdeberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	encoder	zephyrus	1	1	86M	29.1 ms	35.28%	5.88%	16.67%	8.69%	10.43%	18.40%	N/A	N/A	8.69%	10.43%	0.00%	N/A	22.48%	3.75%	16.67%	6.12%	7.34%	8.25%	N/A	N/A	6.12%	7.34%	0.00%	N/A	cần sửainfeasible	–
–	e2_videberta_cross_ce_real_frozenemb_lr1e5	–	argmax	v1	encoder	zephyrus	1	1	86M	25.7 ms	35.28%	5.88%	16.67%	8.69%	10.43%	18.40%	N/A	N/A	8.69%	10.43%	0.00%	N/A	22.48%	3.75%	16.67%	6.12%	7.34%	8.25%	N/A	N/A	6.12%	7.34%	0.00%	N/A	cần sửainfeasible	–
–	e2_videberta_cross_ce_real_frozenemb_lr2e5	–	argmax	v1	encoder	zephyrus	1	1	86M	26.8 ms	35.28%	5.88%	16.67%	8.69%	10.43%	18.40%	N/A	N/A	8.69%	10.43%	0.00%	N/A	22.48%	3.75%	16.67%	6.12%	7.34%	8.25%	N/A	N/A	6.12%	7.34%	0.00%	N/A	cần sửainfeasible	–
–	e2_videberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	encoder		0 1 lỗi	0	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	failed	–
–	e2_videberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	encoder	zephyrus	1	1	86M	31.6 ms	35.28%	5.88%	16.67%	8.69%	10.43%	18.40%	N/A	N/A	8.69%	10.43%	0.00%	N/A	22.48%	3.75%	16.67%	6.12%	7.34%	8.25%	N/A	N/A	6.12%	7.34%	0.00%	N/A	cần sửainfeasible	–
1	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_lv21	jhu-clsp/mmBERT-base	0.076	v2_1	encoder	zephyrus	3	3	111M	42.4 ms	97.50% ±0.32	96.85% ±0.58	97.20% ±0.28	96.97% ±0.34	96.94% ±0.32	97.53% ±0.31	96.83% ±0.69	87.08% ±2.27	97.62% ±0.30	97.61% ±0.33	93.52% ±2.39	95.54% ±1.05	92.85% ±0.46	84.03% ±0.95	87.49% ±2.23	85.12% ±1.37	87.24% ±0.80	93.10% ±0.38	82.35% ±1.31	60.53% ±3.63	85.12% ±0.48	88.55% ±0.13	74.95% ±2.30	75.96% ±1.21	okp85_not_met	–
2	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_lv21	vinai/phobert-base-v2	0.033	v2_1	encoder	zephyrus	1	1	86M	14.7 ms	97.23%	97.07%	97.35%	97.19%	96.95%	97.23%	96.83%	92.42%	96.91%	96.45%	90.08%	96.19%	92.90%	82.51%	84.82%	83.31%	86.64%	93.05%	81.70%	65.10%	83.39%	87.65%	74.51%	73.55%	okp85_not_met	–

Precision theo từng intent (dev và test_full)

STT	Model_Name	Backbone	Threshold	Data	dev Affirm_Pre	dev Negate_Pre	dev Fallback_Pre	dev Callback_Pre	dev Relation_Pre	dev Repeat_Pre	full Affirm_Pre	full Negate_Pre	full Fallback_Pre	full Callback_Pre	full Relation_Pre	full Repeat_Pre
0	e0_prod_sft_hard_none_qc_v16_m1 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	m1	N/A	N/A	N/A	N/A	N/A	N/A	97.82%	89.07%	86.61%	96.12%	94.12%	70.59%
0	e0_prod_sft_hard_none_qc_v16 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v1	N/A	N/A	N/A	N/A	N/A	N/A	97.42%	88.75%	86.67%	95.45%	94.12%	60.00%
0	e0_prod_sft_hard_none_qc_v16_lv2 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2	N/A	N/A	N/A	N/A	N/A	N/A	97.36%	89.38%	81.67%	93.64%	96.47%	70.00%
0	e0_prod_sft_hard_none_qc_v16_lv21 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2_1	N/A	N/A	N/A	N/A	N/A	N/A	97.42%	88.75%	86.67%	95.45%	94.12%	60.00%
1	e3_ensemble_devtop1_familygeomean_m1	–	0.051	m1	98.57%	99.46%	90.55%	100.00%	100.00%	100.00%	99.00%	95.12%	53.73%	95.10%	93.10%	81.82%
2	e3_ensemble_devtop1_familygeomean_tempdev_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	98.29%	99.46%	89.89%	100.00%	100.00%	100.00%	99.01%	95.14%	55.24%	95.15%	93.10%	81.82%
3	e3_ensemble_devtop1_familygeomean_temphard_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	98.29%	99.46%	89.89%	100.00%	100.00%	100.00%	99.01%	95.14%	55.24%	95.15%	93.10%	81.82%
4	e3_ensemble_devtop1_familymean_m1	–	0.112	m1	98.29%	99.24%	91.21%	100.00%	100.00%	100.00%	98.76%	95.32%	56.20%	95.15%	94.25%	83.33%
5	e3_ensemble_sweep_k2_m1	–	0.110	m1	98.03%	99.46%	92.57%	100.00%	100.00%	100.00%	98.24%	95.11%	68.04%	91.51%	96.43%	76.92%
6	e3_ensemble_sweep_k3_m1	–	0.051	m1	98.57%	99.67%	89.57%	100.00%	100.00%	100.00%	98.65%	95.38%	59.56%	91.59%	95.24%	69.23%
7	e3_ensemble_sweep_k4_m1	–	0.058	m1	98.57%	99.67%	89.89%	100.00%	100.00%	100.00%	98.77%	95.37%	59.29%	93.33%	95.29%	69.23%
8	e3_ensemble_sweep_k5_m1	–	0.111	m1	98.45%	99.68%	92.91%	100.00%	100.00%	100.00%	98.72%	95.47%	66.83%	92.52%	95.29%	83.33%
9	e3_ensemble_sweep_k6_m1	–	0.087	m1	98.45%	99.68%	92.91%	100.00%	100.00%	100.00%	98.78%	95.61%	63.38%	92.52%	95.29%	90.91%
10	e3_ensemble_sweep_k7_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.087	m1	98.45%	99.68%	92.91%	100.00%	100.00%	100.00%	98.78%	95.61%	63.38%	92.52%	95.29%	90.91%
11	e3_ensemble_sweep_k8_m1	–	0.087	m1	98.45%	99.68%	92.91%	100.00%	100.00%	100.00%	98.78%	95.61%	63.38%	92.52%	95.29%	90.91%
12	e3_ensemble_top_geomean_m1	–	0.073	m1	98.58%	99.35%	91.21%	100.00%	100.00%	100.00%	98.88%	94.66%	58.52%	93.33%	95.40%	71.43%
13	e3_ensemble_top_mean_m1	–	0.073	m1	98.56%	99.35%	87.37%	100.00%	100.00%	100.00%	99.00%	95.08%	50.75%	95.15%	96.51%	71.43%
14	e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	98.18% ±0.16	99.17% ±0.12	88.85% ±2.88	99.75% ±0.44	99.90% ±0.17	99.47% ±0.92	98.39% ±0.23	94.61% ±0.91	52.03% ±7.64	92.70% ±2.63	94.03% ±0.93	76.39% ±10.49
15	e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1	intfloat/multilingual-e5-base	0.017	m1	98.84%	98.50%	88.26%	100.00%	100.00%	98.41%	98.53%	93.50%	47.48%	89.91%	91.46%	56.25%
16	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m1	jhu-clsp/mmBERT-base	0.155	m1	98.15%	99.03%	91.51%	100.00%	99.70%	100.00%	97.46%	94.21%	60.40%	90.38%	92.77%	66.67%
17	e2_qwen3_17b_base_genlabel_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.081	m1	97.57%	99.24%	85.52%	100.00%	100.00%	100.00%	98.81%	94.79%	50.79%	91.59%	91.95%	68.75%
18	e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-1.7B-Base	0.105	m1	97.62% ±0.34	99.24% ±0.11	87.23% ±0.98	100.00% ±0.00	99.90% ±0.17	99.46% ±0.93	98.77% ±0.14	95.34% ±0.44	51.85% ±0.62	92.46% ±2.62	93.23% ±0.94	80.85% ±6.51
19	e3_ensemble_top_greedy_m1	–	0.113	m1	98.71%	99.25%	90.51%	100.00%	100.00%	100.00%	98.94%	94.99%	54.96%	94.23%	96.47%	73.33%
20	e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m1	intfloat/multilingual-e5-large-instruct	0.017	m1	98.28% ±0.29	99.35% ±0.29	88.30% ±2.78	100.00% ±0.00	100.00% ±0.00	98.96% ±1.80	98.30% ±0.21	94.59% ±0.28	56.57% ±7.19	92.65% ±0.54	94.35% ±0.73	69.06% ±9.28
21	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1	jhu-clsp/mmBERT-base	0.108	m1	98.23% ±0.45	99.24% ±0.39	85.60% ±0.46	99.75% ±0.44	99.90% ±0.17	100.00% ±0.00	98.22% ±0.33	94.34% ±1.09	46.49% ±6.91	90.23% ±0.97	90.99% ±2.11	73.53% ±2.55
22	e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1	Qwen/Qwen3-4B-Base	0.056	m1	98.23% ±0.41	99.71% ±0.17	86.91% ±0.75	100.00% ±0.00	99.61% ±0.17	100.00% ±0.00	98.64% ±0.24	95.61% ±0.53	54.24% ±2.06	94.53% ±1.38	93.68% ±0.61	64.48% ±2.09
23	e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1	unsloth/gemma-3-1b-pt	0.045	m1	98.55% ±0.41	98.92% ±0.00	86.40% ±1.31	100.00% ±0.00	100.00% ±0.00	100.00% ±0.00	98.66% ±0.21	94.55% ±0.01	50.97% ±1.37	94.56% ±1.92	93.07% ±1.07	79.62% ±14.69
24	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepeval_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1, nhãn m1	vinai/phobert-base-v2	0.033	m1	98.42% ±0.37	99.17% ±0.16	85.98% ±0.46	100.00% ±0.00	99.51% ±0.17	100.00% ±0.00	98.37% ±0.27	94.16% ±0.28	49.21% ±2.29	92.19% ±2.52	90.86% ±1.26	72.88% ±8.39
25	e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1	unsloth/gemma-3-1b-pt	0.045	m1	98.51% ±0.58	99.24% ±0.43	86.20% ±1.82	100.00% ±0.00	99.90% ±0.17	99.47% ±0.92	98.45% ±0.14	94.07% ±1.04	48.95% ±5.81	95.66% ±2.00	90.41% ±2.81	70.04% ±11.55
26	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1	jhu-clsp/mmBERT-base	0.052	m1	98.71% ±0.30	99.24% ±0.22	87.51% ±1.41	100.00% ±0.00	99.61% ±0.45	98.94% ±0.92	98.17% ±0.12	94.11% ±0.27	51.97% ±3.78	91.02% ±0.78	93.22% ±1.49	59.53% ±10.40
27	e2_mmbert_base_cross_ce_synth1x_frozenemb_lr2e5_fbmatched_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	jhu-clsp/mmBERT-base	0.051	m1	98.20% ±0.45	99.35% ±0.19	88.97% ±1.09	100.00% ±0.00	99.90% ±0.17	99.48% ±0.90	98.33% ±0.15	94.74% ±0.28	53.74% ±4.22	94.39% ±1.42	93.33% ±2.07	70.48% ±3.43
28	e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.118	m1	98.14% ±0.72	99.17% ±0.06	88.56% ±1.42	100.00% ±0.00	99.71% ±0.29	100.00% ±0.00	98.39% ±0.46	94.48% ±0.06	52.67% ±6.69	93.76% ±1.46	91.88% ±0.84	70.09% ±8.92
29	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1	vinai/phobert-base-v2	0.008	m1	98.30%	99.24%	89.17%	100.00%	99.71%	100.00%	98.27%	93.65%	53.91%	91.43%	89.77%	100.00%
30	e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1	Qwen/Qwen3-4B-Base	0.162	m1	97.59%	99.56%	87.90%	100.00%	99.71%	100.00%	98.76%	95.66%	51.92%	93.40%	89.47%	66.67%
31	e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1	Qwen/Qwen3-8B-Base	0.170	m1	98.45%	99.57%	91.14%	100.00%	99.41%	100.00%	98.53%	95.73%	61.61%	93.33%	89.47%	68.75%
32	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_m1	vinai/phobert-base-v2	0.073	m1	98.15% ±0.23	99.35% ±0.11	88.67% ±1.86	100.00% ±0.00	99.32% ±0.16	99.46% ±0.93	98.17% ±0.13	95.56% ±0.41	53.37% ±1.20	92.74% ±1.28	88.78% ±5.35	69.07% ±9.53
33	e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	98.52% ±0.21	99.35% ±0.00	87.50% ±1.33	100.00% ±0.00	100.00% ±0.00	100.00% ±0.00	98.37% ±0.55	93.69% ±0.49	54.56% ±1.46	94.15% ±1.83	92.68% ±0.81	67.86% ±6.19
34	e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-base	0.031	m1	98.31% ±0.32	98.41% ±0.31	86.29% ±1.98	99.74% ±0.44	99.90% ±0.17	100.00% ±0.00	98.58% ±0.06	93.74% ±0.49	52.58% ±5.12	89.85% ±0.83	91.76% ±0.26	69.70% ±5.25
35	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.061	m1	98.51% ±0.23	98.85% ±0.67	88.13% ±1.98	100.00% ±0.00	99.80% ±0.34	99.47% ±0.92	98.11% ±0.39	94.48% ±1.02	50.88% ±5.44	94.80% ±2.89	92.41% ±0.27	68.81% ±7.84
36	e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1	Qwen/Qwen3-4B-Base	0.131	m1	97.85% ±0.38	99.49% ±0.06	87.58% ±3.37	100.00% ±0.00	99.80% ±0.17	100.00% ±0.00	98.61% ±0.11	94.99% ±0.41	60.01% ±8.76	93.15% ±3.55	93.62% ±2.18	69.47% ±18.94
37	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.043	m1	97.67% ±0.66	98.73% ±0.33	86.69% ±2.16	99.87% ±0.22	99.90% ±0.17	100.00% ±0.00	98.20% ±0.16	93.52% ±0.43	52.04% ±2.31	93.03% ±1.80	92.72% ±1.50	77.78% ±19.25
38	e2_gemma3_1b_pt_cross_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.053	m1	98.41% ±0.15	99.20% ±0.27	86.25% ±0.75	100.00% ±0.00	99.90% ±0.17	98.94% ±0.92	98.30% ±0.24	94.12% ±0.46	53.12% ±2.91	94.39% ±1.39	92.54% ±1.20	71.78% ±9.04
39	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1	vinai/phobert-base-v2	0.037	m1	98.32% ±0.33	99.06% ±0.16	87.05% ±0.99	100.00% ±0.00	99.71% ±0.51	100.00% ±0.00	98.26% ±0.09	93.76% ±1.31	53.37% ±2.16	91.49% ±2.29	93.10% ±0.14	83.42% ±14.47
40	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m1	vinai/phobert-base-v2	0.085	m1	97.96% ±0.06	99.17% ±0.69	89.73% ±2.51	100.00% ±0.00	99.51% ±0.17	100.00% ±0.00	97.96% ±0.42	94.62% ±0.31	58.16% ±1.84	90.46% ±0.48	90.83% ±4.03	74.94% ±7.54
41	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_m1	vinai/phobert-base-v2	0.028	m1	98.18% ±0.42	99.17% ±0.25	86.64% ±1.02	100.00% ±0.00	99.70% ±0.51	100.00% ±0.00	98.40% ±0.18	94.07% ±1.24	50.66% ±1.25	94.18% ±1.55	92.95% ±1.14	60.00% ±9.21
42	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1	jhu-clsp/mmBERT-base	0.082	m1	98.14% ±0.14	99.24% ±0.11	87.82% ±2.47	100.00% ±0.00	99.80% ±0.17	99.47% ±0.92	98.18% ±0.20	93.83% ±0.76	53.43% ±7.60	92.86% ±1.31	90.90% ±2.24	59.55% ±4.41
43	e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.057	m1	97.86% ±0.47	99.28% ±0.40	87.26% ±1.17	100.00% ±0.00	99.71% ±0.29	100.00% ±0.00	98.46% ±0.09	94.67% ±0.55	48.63% ±1.08	94.09% ±1.73	92.54% ±2.51	62.08% ±6.88
44	e2_qwen3_17b_base_cross_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.101	m1	98.17% ±0.20	98.99% ±0.12	87.09% ±2.39	99.62% ±0.00	99.71% ±0.00	100.00% ±0.00	98.13% ±0.36	94.53% ±0.64	51.90% ±4.14	93.88% ±1.97	92.84% ±0.47	64.35% ±4.38
45	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m1	vinai/phobert-base-v2	0.052	m1	98.14% ±0.60	98.99% ±0.61	86.32% ±0.33	100.00% ±0.00	99.90% ±0.17	99.47% ±0.92	98.40% ±0.28	93.75% ±2.31	52.81% ±3.88	93.28% ±0.87	92.49% ±2.81	76.79% ±10.67
46	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1	vinai/phobert-base-v2	0.077	m1	97.58% ±0.62	99.06% ±0.41	87.38% ±2.34	100.00% ±0.00	99.61% ±0.45	100.00% ±0.00	98.31% ±0.14	94.60% ±0.53	55.41% ±7.56	91.58% ±1.45	92.26% ±1.59	88.60% ±11.54
47	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	vinai/phobert-base-v2	0.058	m1	97.85% ±0.37	98.56% ±0.27	86.74% ±1.36	100.00% ±0.00	99.80% ±0.17	99.48% ±0.90	98.28% ±0.30	92.98% ±0.63	54.17% ±2.40	93.26% ±1.72	90.66% ±2.40	62.66% ±8.27
48	e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1	unsloth/gemma-3-1b-pt	0.056	m1	98.51% ±0.08	99.10% ±0.27	87.03% ±0.96	100.00% ±0.00	99.90% ±0.17	99.48% ±0.90	98.37% ±0.13	94.52% ±1.16	51.31% ±3.86	91.71% ±1.28	93.58% ±0.88	70.69% ±11.52
49	e2_gemma3_270m_cross_ce_real_lora_m1gemma-3-270m · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn m1	unsloth/gemma-3-270m	0.098	m1	97.25% ±0.75	98.81% ±0.28	88.34% ±2.79	100.00% ±0.00	100.00% ±0.00	98.97% ±1.78	97.59% ±0.64	93.72% ±0.95	49.34% ±2.36	91.03% ±0.48	93.31% ±1.73	55.96% ±5.27
50	e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-large-instruct	0.029	m1	98.10% ±0.08	99.06% ±0.31	89.53% ±3.05	99.87% ±0.22	99.90% ±0.17	99.47% ±0.92	98.39% ±0.21	94.56% ±0.83	54.55% ±1.30	93.62% ±1.40	93.94% ±1.09	71.74% ±8.79
51	e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.034	m1	98.41% ±0.14	98.88% ±0.23	86.01% ±0.61	99.87% ±0.22	99.80% ±0.34	100.00% ±0.00	98.56% ±0.22	93.97% ±0.45	48.31% ±2.61	92.44% ±3.48	92.87% ±1.79	68.74% ±9.43
52	e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.051	m1	98.18% ±0.16	98.92% ±0.32	86.93% ±0.95	100.00% ±0.00	99.41% ±0.50	100.00% ±0.00	98.22% ±0.04	94.06% ±1.13	50.66% ±0.98	94.94% ±1.46	91.59% ±1.29	60.73% ±15.51
53	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1	jhu-clsp/mmBERT-base	0.040	m1	98.52% ±0.30	98.92% ±0.32	87.16% ±1.42	100.00% ±0.00	99.61% ±0.44	99.47% ±0.92	98.37% ±0.36	94.68% ±0.42	48.49% ±3.91	93.80% ±1.99	92.23% ±3.02	62.95% ±8.76
54	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1	jhu-clsp/mmBERT-base	0.057	m1	98.13% ±0.37	98.95% ±0.27	85.72% ±0.22	100.00% ±0.00	99.61% ±0.34	99.47% ±0.92	98.38% ±0.24	93.92% ±0.51	51.69% ±4.25	95.03% ±1.02	93.19% ±0.80	62.70% ±11.25
55	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.070	m1	98.43%	98.82%	90.11%	100.00%	100.00%	100.00%	98.09%	94.63%	56.83%	90.57%	92.77%	64.71%
56	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.047	m1	98.34% ±0.79	99.02% ±0.65	88.75% ±2.12	100.00% ±0.00	99.80% ±0.17	99.46% ±0.93	98.11% ±0.21	94.11% ±0.38	51.25% ±1.85	92.91% ±2.13	93.20% ±0.58	68.86% ±14.42
57	e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1	Qwen/Qwen3-1.7B-Base	0.095	m1	97.15% ±0.66	99.20% ±0.27	89.03% ±1.08	100.00% ±0.00	100.00% ±0.00	100.00% ±0.00	98.50% ±0.35	94.61% ±0.67	54.24% ±4.13	96.33% ±0.53	93.93% ±2.16	83.60% ±5.16
58	e2_xlmr_base_khonly_ce_real_frozenemb_m1XLM-R base · chỉ câu KH · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1	FacebookAI/xlm-roberta-base	0.012	m1	97.40%	98.06%	86.32%	100.00%	99.71%	100.00%	97.81%	94.40%	50.96%	86.73%	92.05%	61.54%
59	e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.042	m1	97.81% ±0.35	99.17% ±0.06	86.01% ±1.25	100.00% ±0.00	100.00% ±0.00	99.46% ±0.93	98.07% ±0.15	94.88% ±0.65	50.30% ±2.58	92.91% ±2.57	91.95% ±1.73	78.46% ±3.41
60	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m1	vinai/phobert-base-v2	0.024	m1	97.71% ±0.36	99.24% ±0.29	85.50% ±0.43	100.00% ±0.00	99.80% ±0.17	100.00% ±0.00	98.40% ±0.46	94.23% ±0.18	49.76% ±0.57	92.52% ±1.48	92.97% ±1.23	81.94% ±6.21
61	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m1	vinai/phobert-base-v2	0.031	m1	97.50% ±0.16	99.20% ±0.13	87.76% ±1.47	100.00% ±0.00	99.70% ±0.29	100.00% ±0.00	97.91% ±0.28	95.03% ±0.17	52.21% ±5.20	92.71% ±1.74	92.30% ±2.77	77.78% ±3.85
62	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m1	vinai/phobert-base-v2	0.031	m1	97.76% ±0.29	98.95% ±0.25	86.01% ±1.30	99.87% ±0.22	99.80% ±0.17	100.00% ±0.00	98.26% ±0.27	94.13% ±0.35	53.67% ±4.36	91.21% ±1.21	92.49% ±1.75	79.58% ±2.47
63	e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1	unsloth/gemma-3-1b-pt	0.096	m1	98.00% ±0.48	99.13% ±0.28	87.85% ±1.14	100.00% ±0.00	99.90% ±0.17	98.96% ±1.80	98.24% ±0.17	93.18% ±0.45	53.52% ±5.73	94.13% ±2.45	92.37% ±1.28	45.76% ±16.77
64	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1	jhu-clsp/mmBERT-base	0.061	m1	98.17% ±0.38	98.84% ±0.23	85.81% ±1.20	100.00% ±0.00	99.90% ±0.17	99.47% ±0.92	98.27% ±0.55	94.34% ±0.26	46.75% ±2.48	93.00% ±3.38	92.52% ±1.61	66.50% ±18.84
65	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.017	m1	97.83% ±0.45	99.17% ±0.33	89.67% ±1.00	100.00% ±0.00	99.71% ±0.29	99.46% ±0.93	98.09% ±0.06	94.37% ±0.12	54.26% ±1.37	94.16% ±0.89	91.27% ±2.42	88.43% ±11.14
66	e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1	vinai/phobert-base-v2	0.026	m1	97.74% ±0.50	98.88% ±0.17	91.24% ±1.99	100.00% ±0.00	99.80% ±0.17	98.42% ±1.56	98.11% ±0.24	93.64% ±0.72	57.88% ±5.62	92.71% ±1.39	91.22% ±2.18	80.56% ±7.35
67	e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.065	m1	97.79% ±0.22	99.10% ±0.25	86.29% ±0.83	100.00% ±0.00	99.61% ±0.17	99.46% ±0.93	98.14% ±0.37	93.97% ±0.35	53.30% ±2.55	91.48% ±1.71	92.64% ±1.28	69.49% ±11.18
68	e2_xlmr_base_cross_logitadjust_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.018	m1	97.75% ±0.07	98.77% ±0.40	86.18% ±0.47	99.62% ±0.66	99.71% ±0.29	99.47% ±0.92	98.16% ±0.03	94.55% ±0.36	47.71% ±2.34	89.68% ±2.53	89.43% ±3.49	64.29% ±0.00
69	e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-large-instruct	0.026	m1	98.33% ±0.23	98.89% ±0.16	89.67% ±2.24	99.87% ±0.22	100.00% ±0.00	99.46% ±0.93	98.33% ±0.18	94.38% ±0.28	58.15% ±6.44	92.97% ±1.93	94.83% ±0.62	65.28% ±2.41
70	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1	jhu-clsp/mmBERT-base	0.072	m1	98.16% ±0.60	98.77% ±0.35	89.99% ±0.92	100.00% ±0.00	99.80% ±0.17	100.00% ±0.00	97.85% ±0.33	94.04% ±0.47	54.46% ±3.67	93.53% ±2.42	92.57% ±1.59	64.66% ±9.87
71	e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1	jhu-clsp/mmBERT-base	0.082	m1	98.04% ±0.57	98.85% ±0.55	87.19% ±0.22	100.00% ±0.00	99.51% ±0.34	99.47% ±0.92	97.94% ±0.70	94.52% ±0.65	53.22% ±1.42	94.95% ±0.84	93.08% ±1.89	60.83% ±6.82
72	e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.067	m1	97.42%	98.92%	86.57%	100.00%	100.00%	100.00%	98.11%	94.97%	45.61%	93.27%	94.05%	66.67%
73	e2_visobert_cross_ce_real_frozenemb_lr4e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	uitnlp/visobert	0.037	m1	98.69%	98.07%	88.13%	100.00%	99.41%	98.39%	98.58%	93.20%	40.06%	91.18%	91.76%	69.23%
74	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1	FacebookAI/xlm-roberta-base	0.027	m1	97.70%	98.49%	87.19%	99.62%	100.00%	100.00%	97.91%	93.32%	47.69%	90.74%	91.86%	69.23%
75	e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m1mmBERT base · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.057	m1	98.12% ±0.01	98.46% ±0.79	87.33% ±1.73	100.00% ±0.00	99.80% ±0.17	98.99% ±1.75	97.99% ±0.08	94.21% ±0.76	51.09% ±4.91	96.64% ±1.47	97.94% ±1.90	83.80% ±7.65
76	e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.206	m1	97.64% ±0.09	98.67% ±0.06	91.15% ±2.87	100.00% ±0.00	99.61% ±0.45	100.00% ±0.00	97.53% ±0.44	93.94% ±0.06	65.69% ±12.86	91.88% ±2.08	91.74% ±1.61	67.97% ±15.15
77	e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.034	m1	97.62% ±0.17	98.59% ±0.11	87.21% ±1.58	100.00% ±0.00	100.00% ±0.00	99.46% ±0.93	98.05% ±0.03	93.87% ±0.63	57.67% ±3.42	91.62% ±0.86	90.43% ±1.45	74.77% ±12.27
78	e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.052	m1	97.40% ±0.15	98.92% ±0.32	88.38% ±2.55	100.00% ±0.00	99.90% ±0.17	99.47% ±0.92	97.80% ±0.20	94.61% ±0.67	56.34% ±6.39	93.32% ±1.57	92.01% ±0.86	72.38% ±13.20
79	e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.033	m1	97.34% ±0.30	98.91% ±0.39	87.00% ±2.02	100.00% ±0.00	99.80% ±0.17	100.00% ±0.00	97.72% ±0.29	94.54% ±0.57	55.68% ±4.29	92.42% ±2.30	92.34% ±0.47	75.84% ±12.65
80	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1	jhu-clsp/mmBERT-base	0.144	m1	97.86% ±0.84	98.81% ±0.57	88.49% ±2.39	100.00% ±0.00	99.41% ±0.29	100.00% ±0.00	98.05% ±0.29	94.03% ±0.06	51.64% ±3.49	94.66% ±2.89	91.51% ±0.54	64.48% ±2.09
81	e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.082	m1	98.19% ±0.33	98.95% ±0.38	88.91% ±3.07	100.00% ±0.00	99.80% ±0.17	99.47% ±0.92	98.21% ±0.23	94.61% ±0.82	52.31% ±7.01	91.34% ±2.08	92.06% ±0.66	63.18% ±14.57
82	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1	jhu-clsp/mmBERT-base	0.108	m1	97.78% ±0.64	99.13% ±0.11	87.74% ±2.32	100.00% ±0.00	99.71% ±0.29	100.00% ±0.00	97.96% ±0.59	94.07% ±0.72	58.10% ±5.14	94.47% ±3.03	92.01% ±1.46	66.07% ±4.72
83	e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-0.6B-Base	0.075	m1	97.76% ±0.16	98.77% ±0.16	87.71% ±2.13	100.00% ±0.00	99.90% ±0.17	97.46% ±1.68	98.31% ±0.23	94.95% ±0.62	50.70% ±6.26	91.27% ±1.79	95.39% ±1.98	63.43% ±2.89
84	e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1	Qwen/Qwen3-0.6B-Base	0.077	m1	96.88% ±0.14	98.45% ±0.34	86.78% ±2.11	100.00% ±0.00	99.80% ±0.34	100.00% ±0.00	97.46% ±0.34	93.26% ±0.27	45.69% ±4.25	95.02% ±1.00	94.90% ±0.64	73.81% ±8.58
85	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_m1	vinai/phobert-base-v2	0.093	m1	97.80% ±0.54	98.63% ±0.58	87.05% ±1.23	99.24% ±0.75	99.51% ±0.34	99.48% ±0.90	98.53% ±0.36	93.92% ±1.66	55.88% ±1.25	90.86% ±2.82	91.59% ±2.62	70.19% ±10.98
86	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m1	vinai/phobert-base-v2	0.104	m1	97.59% ±0.49	99.09% ±0.16	88.27% ±2.74	99.74% ±0.22	99.61% ±0.17	100.00% ±0.00	97.90% ±0.56	93.59% ±0.42	58.77% ±8.67	91.60% ±1.54	92.01% ±1.79	81.20% ±17.07
87	e2_xlmr_base_cross_ce_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.044	m1	97.25% ±0.56	98.91% ±0.22	87.08% ±2.15	99.75% ±0.22	99.90% ±0.17	99.45% ±0.95	97.93% ±0.41	93.84% ±0.55	50.55% ±7.61	89.48% ±0.43	90.04% ±1.89	69.80% ±3.04
88	e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.019	m1	97.48% ±0.52	98.88% ±0.25	87.70% ±1.98	99.75% ±0.22	99.90% ±0.17	98.45% ±1.54	97.83% ±0.48	93.96% ±0.57	52.15% ±5.51	89.88% ±0.78	90.26% ±2.18	66.83% ±3.75
89	e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.049	m1	97.65% ±0.34	98.70% ±0.39	85.91% ±0.51	99.25% ±1.30	99.80% ±0.17	100.00% ±0.00	98.14% ±0.63	94.08% ±0.43	52.29% ±0.98	93.01% ±1.81	91.57% ±1.33	61.57% ±5.61
90	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.159	m1	97.15% ±0.36	98.48% ±0.31	86.91% ±0.81	99.12% ±1.21	99.80% ±0.17	98.37% ±1.61	97.65% ±0.33	93.46% ±0.79	56.95% ±8.27	91.88% ±1.06	92.56% ±2.00	65.64% ±9.77
91	e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-4B	0.061	m1	95.41% ±2.81	97.93% ±0.80	85.39% ±0.10	99.81% ±0.28	99.26% ±0.62	99.22% ±1.10	96.53% ±0.70	92.69% ±0.22	48.84% ±1.10	98.28% ±0.94	90.98% ±5.12	67.92% ±31.23
92	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.025	m1	97.74%	98.58%	86.74%	99.62%	99.41%	100.00%	97.62%	93.93%	55.02%	91.51%	88.37%	50.00%
93	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.152	m1	95.86% ±0.82	98.47% ±0.36	88.02% ±1.88	99.74% ±0.22	99.41% ±0.50	98.42% ±0.01	96.91% ±0.65	91.65% ±1.17	53.08% ±4.73	91.49% ±1.38	90.28% ±1.80	58.75% ±5.34
94	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m1	vinai/phobert-base-v2	0.035	m1	97.61% ±0.29	98.99% ±0.12	86.72% ±1.78	99.13% ±1.51	99.32% ±0.73	99.45% ±0.95	98.30% ±0.11	94.27% ±0.95	53.77% ±5.56	91.83% ±1.02	91.11% ±3.24	70.83% ±7.22
95	e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-base	0.052	m1	97.24%	97.95%	86.07%	97.75%	100.00%	98.36%	97.69%	93.63%	48.85%	87.27%	95.00%	50.00%
96	e2_visobert_cross_ce_real_frozenemb_lr2e5_m1	uitnlp/visobert	0.027	m1	97.84%	97.32%	87.59%	100.00%	99.41%	98.36%	97.72%	92.12%	42.47%	93.07%	90.91%	58.33%
97	e2_visobert_cross_ce_real_frozenemb_lr1e5_m1	uitnlp/visobert	0.186	m1	96.66%	96.26%	86.81%	98.86%	99.12%	100.00%	97.23%	90.52%	44.69%	92.16%	88.75%	57.14%
98	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1	FacebookAI/xlm-roberta-base	0.161	m1	96.31%	98.15%	86.81%	97.75%	100.00%	100.00%	96.97%	93.35%	52.72%	87.39%	91.46%	46.67%
99	e1_tfidf_hybrid_qwen3emb_lr_m1TF-IDF câu KH (char 1-5, word 1-2-gram, giữ từ 1 ký tự, thêm kênh char bỏ dấu) + câu bot (char 2-4) + tương tác loại câu bot × từ câu KH · vector đóng băng Qwen3-Embedding-0.6B (cặp bot + KH, pooling last_token), trọng số dense 1 · LogReg · train thật, nhãn m1	Qwen/Qwen3-Embedding-0.6B	0.149	m1	96.26%	96.36%	85.77%	99.61%	99.70%	100.00%	96.57%	90.24%	47.06%	92.71%	93.59%	72.73%
100	e1_tfidf_svmcal_ce_synth1x_fbmatched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	TF-IDF	0.207	m1	95.57%	95.96%	86.84%	99.23%	99.70%	100.00%	96.36%	90.07%	57.43%	91.26%	94.81%	80.00%
101	e1_tfidf_svmcal_ce_synth1x_nofb_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent omi_intent_repeat, callback, relationship, negate, affirm), nhãn m1	TF-IDF	0.164	m1	95.82%	95.95%	85.87%	98.86%	99.70%	100.00%	96.53%	90.38%	55.02%	91.26%	94.81%	72.73%
102	e1_tfidf_svmcal_ce_synth1x_noqparticle_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm q_particle), nhãn m1	TF-IDF	0.280	m1	95.55%	95.65%	85.82%	99.61%	99.70%	100.00%	96.95%	90.38%	47.24%	90.38%	96.00%	76.92%
103	e1_tfidf_svmcal_ce_real_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn m1	TF-IDF	0.178	m1	95.69%	95.86%	86.42%	99.24%	99.70%	100.00%	97.23%	89.13%	54.84%	91.26%	94.81%	72.73%
104	e1_tfidf_svmcal_ce_synth1x_nofiller_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm filler_only), nhãn m1	TF-IDF	0.267	m1	95.43%	96.05%	85.13%	99.62%	99.70%	100.00%	96.96%	91.00%	47.51%	90.38%	95.95%	81.82%
105	e1_tfidf_svmcal_ce_synth1x_noofftopic_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm prefix_offtopic), nhãn m1	TF-IDF	0.298	m1	95.41%	95.75%	85.45%	100.00%	99.70%	100.00%	96.83%	90.58%	48.58%	90.38%	94.81%	71.43%
106	e2_dual_qwen3emb06b_concat_ce_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.227	m1	93.92% ±0.94	97.52% ±0.11	86.84% ±0.93	97.73% ±0.37	98.22% ±0.77	99.47% ±0.92	95.10% ±0.80	90.20% ±0.66	46.30% ±3.09	95.13% ±0.46	86.12% ±1.86	92.96% ±6.12
107	e1_tfidf_svmcal_ce_synth1x_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1, nhãn m1	TF-IDF	0.282	m1	95.28%	96.06%	85.39%	99.61%	99.70%	100.00%	96.83%	90.50%	47.04%	90.38%	94.74%	71.43%
108	e1_tfidf_svmcal_ce_synth1x_matched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ intent của train thật), nhãn m1	TF-IDF	0.190	m1	95.68%	95.66%	86.04%	98.85%	99.70%	100.00%	97.11%	88.97%	55.14%	91.26%	94.74%	72.73%
109	e1_tfidf_svmcal_ce_synth1x_nomoney_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm money), nhãn m1	TF-IDF	0.310	m1	95.16%	95.85%	86.69%	99.61%	99.70%	100.00%	96.90%	90.42%	49.39%	90.38%	94.87%	76.92%
110	e1_tfidf_svmcal_ce_synth1x_fbonly_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent nlu_fallback), nhãn m1	TF-IDF	0.356	m1	95.71%	95.57%	87.98%	99.62%	99.70%	100.00%	97.10%	90.42%	44.65%	90.38%	94.81%	71.43%
–	e2_bamibert_cross_ce_real_frozenemb_lr1e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_bamibert_cross_ce_real_frozenemb_lr2e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_bamibert_cross_ce_real_frozenemb_lr4e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e3_fbverifier_devtop1_familygeomean_fbf1_m1	–	hard	m1	98.04%	98.93%	94.84%	100.00%	100.00%	96.92%	98.13%	94.37%	81.21%	91.67%	93.18%	64.71%
–	e3_fbverifier_devtop1_familygeomean_macrof1_m1	–	hard	m1	98.04%	98.93%	94.84%	100.00%	100.00%	96.92%	98.13%	94.37%	81.21%	91.67%	93.18%	64.71%
1	e3_ensemble_mean_all	–	0.065	v1	99.40%	98.97%	87.50%	98.15%	98.86%	100.00%	97.81%	95.09%	55.87%	94.39%	92.05%	62.50%
2	e3_ensemble_greedy_all	–	0.070	v1	99.15%	99.17%	86.87%	98.13%	98.86%	98.41%	97.69%	95.27%	55.47%	93.52%	91.95%	62.50%
3	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5	–	0.023	v1	97.77%	99.17%	89.55%	98.51%	100.00%	98.39%	97.24%	95.33%	60.55%	90.99%	90.59%	47.62%
4	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1	vinai/phobert-base-v2	0.011	v1	97.87%	99.48%	88.58%	98.14%	98.55%	100.00%	96.72%	94.96%	56.65%	96.19%	92.13%	57.14%
5	e2_xlmr_base_cross_wcesqrt_real_frozenemb	FacebookAI/xlm-roberta-base	0.021	v1	98.47% ±0.50	98.69% ±0.12	85.62% ±0.52	98.39% ±0.21	98.49% ±0.43	100.00% ±0.00	97.47% ±0.12	94.27% ±0.47	56.71% ±2.55	93.88% ±2.19	89.46% ±2.43	60.58% ±5.43
6	e3_prior_em_mmbert_bottype	–	0.082	v1	98.56%	99.17%	86.49%	98.50%	98.86%	100.00%	96.82%	96.35%	62.38%	95.05%	29.34%	66.67%
7	e3_prior_em_mmbert_global	–	0.082	v1	98.56%	99.17%	86.49%	98.50%	98.86%	100.00%	96.05%	96.63%	57.69%	94.29%	92.59%	76.92%
8	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5	jhu-clsp/mmBERT-base	0.061	v1	98.22% ±0.12	99.24% ±0.16	87.26% ±0.40	98.38% ±0.21	98.77% ±1.14	100.00% ±0.00	97.56% ±0.18	95.31% ±0.70	55.89% ±3.08	94.33% ±2.34	90.83% ±2.36	68.01% ±3.01
9	e2_xlmr_base_cross_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	98.36% ±0.28	98.53% ±0.43	89.35% ±2.96	98.27% ±0.42	98.38% ±0.43	97.87% ±2.42	97.31% ±0.22	93.60% ±0.34	60.95% ±0.90	93.82% ±1.00	89.65% ±1.91	63.16% ±7.96
10	e2_xlmr_base_cross_logitadjust_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	98.44% ±0.23	99.07% ±0.55	86.57% ±1.19	98.28% ±0.21	99.05% ±0.16	97.88% ±0.87	97.67% ±0.24	94.49% ±0.42	55.66% ±1.10	94.74% ±1.40	92.46% ±1.68	52.73% ±2.78
11	e2_qwen3_06b_base_cross_ce_real_lora	Qwen/Qwen3-0.6B-Base	0.098	v1	98.32%	97.85%	89.16%	97.77%	99.13%	100.00%	97.39%	91.89%	60.09%	95.24%	90.00%	62.50%
12	e3_ensemble_stack_all	–	0.088	v1	98.43%	98.97%	85.00%	98.49%	99.14%	98.33%	97.38%	93.35%	57.45%	97.03%	92.94%	66.67%
13	e2_qwen3_17b_base_cross_ce_real_lora	Qwen/Qwen3-1.7B-Base	0.083	v1	98.17% ±0.59	98.66% ±0.46	85.65% ±0.98	98.48% ±0.03	98.57% ±0.57	98.91% ±0.95	97.38% ±0.12	94.19% ±0.75	55.59% ±2.51	96.15% ±0.91	92.46% ±0.51	65.28% ±2.41
14	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5	vinai/phobert-base-v2	0.022	v1	97.55% ±0.37	98.56% ±0.18	89.99% ±1.66	97.79% ±0.36	98.66% ±0.58	99.46% ±0.93	97.29% ±0.15	94.25% ±0.16	59.28% ±0.95	95.01% ±2.23	90.69% ±2.26	69.29% ±5.68
15	e2_me5_base_cross_ce_real_frozenemb_lr1e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	intfloat/multilingual-e5-base	0.088	v1	97.84%	98.35%	88.50%	97.79%	98.29%	98.36%	97.32%	93.03%	58.80%	90.83%	91.95%	52.94%
16	e2_me5_base_cross_ce_real_frozenemb_lr4e5	–	0.028	v1	98.91%	98.36%	85.81%	98.50%	98.58%	98.36%	97.35%	93.72%	60.00%	94.39%	89.89%	71.43%
17	e2_xlmr_base_khonly_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.076	v1	97.37% ±0.43	97.74% ±0.37	89.11% ±1.40	98.02% ±0.21	98.38% ±0.43	99.46% ±0.93	96.95% ±0.16	93.61% ±0.73	60.29% ±3.36	92.93% ±0.47	92.08% ±1.08	67.42% ±2.73
18	e2_me5_base_cross_ce_real_frozenemb_lr2e5	intfloat/multilingual-e5-base	0.056	v1	98.63% ±0.42	98.10% ±0.25	88.57% ±0.82	97.91% ±0.55	98.49% ±0.16	99.44% ±0.96	97.75% ±0.17	93.21% ±0.63	59.69% ±1.44	91.57% ±0.94	91.35% ±2.18	63.99% ±7.59
19	e2_gemma3_270m_cross_ce_real_lora	unsloth/gemma-3-270m	0.116	v1	97.80% ±0.13	98.09% ±0.25	87.65% ±2.07	98.25% ±0.20	98.57% ±0.48	99.47% ±0.92	97.35% ±0.33	93.76% ±0.54	56.64% ±2.95	93.69% ±1.34	90.32% ±0.38	57.68% ±4.94
20	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen	Qwen/Qwen3-Embedding-0.6B	0.114	v1	96.91% ±0.35	98.07% ±0.21	86.59% ±0.99	98.38% ±0.21	98.56% ±0.28	100.00% ±0.00	96.92% ±0.05	93.45% ±0.13	56.06% ±0.80	91.72% ±2.09	92.12% ±1.84	61.47% ±5.60
21	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5	–	0.075	v1	97.88%	98.86%	87.46%	98.50%	100.00%	100.00%	97.58%	94.95%	56.78%	96.12%	92.05%	62.50%
22	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5	–	0.026	v1	96.94%	98.34%	87.15%	96.03%	99.13%	100.00%	96.82%	93.37%	56.43%	92.59%	90.91%	100.00%
23	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	FacebookAI/xlm-roberta-base	0.123	v1	96.80%	98.24%	86.55%	97.79%	98.84%	94.64%	96.70%	93.83%	60.00%	91.15%	92.68%	28.57%
24	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5	jhu-clsp/mmBERT-base	0.120	v1	98.44% ±0.17	98.76% ±0.28	86.08% ±0.18	98.12% ±0.52	98.01% ±0.02	97.64% ±1.09	97.34% ±0.08	93.53% ±0.65	60.29% ±0.79	96.16% ±1.31	88.03% ±4.07	54.41% ±6.24
25	e2_visobert_cross_ce_real_frozenemb_lr2e5	–	0.093	v1	96.10%	97.40%	85.86%	97.74%	98.86%	98.36%	97.15%	92.59%	58.18%	94.23%	87.78%	55.56%
26	e2_visobert_cross_ce_real_frozenemb_lr4e5	–	0.066	v1	97.37%	97.44%	87.94%	98.48%	97.48%	98.36%	97.14%	93.22%	55.56%	95.19%	88.89%	64.71%
27	e1_tfidf_svmcal_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn v1	TF-IDF	0.123	v1	95.84%	95.79%	85.17%	97.41%	100.00%	98.41%	97.10%	93.02%	54.40%	93.40%	95.00%	71.43%
28	e1_tfidf_logreg_balanced_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg cân bằng lớp · train thật, nhãn v1	TF-IDF	0.256	v1	95.50%	96.17%	86.62%	97.41%	98.85%	98.44%	96.67%	93.15%	50.00%	89.09%	88.30%	55.00%
29	e1_tfidf_logreg_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật, nhãn v1	TF-IDF	0.125	v1	95.62%	96.18%	86.01%	97.77%	99.13%	98.39%	96.96%	92.50%	50.94%	93.27%	90.36%	66.67%
30	e2_visobert_cross_ce_real_frozenemb_lr1e5	–	0.232	v1	94.27%	97.17%	88.04%	95.64%	97.17%	95.16%	96.63%	92.15%	58.54%	89.91%	87.50%	50.00%
31	e1_tfidf_logreg_ce_synth1x_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật + synth ×1, nhãn v1	TF-IDF	0.234	v1	94.63%	95.98%	85.82%	97.77%	99.70%	98.36%	96.46%	92.23%	54.62%	93.33%	91.46%	62.50%
–	e2_mdeberta_cross_ce_real_frozenemb_lr1e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	N/A	35.28%	N/A	N/A	N/A	N/A	N/A	22.48%	N/A	N/A	N/A	N/A
–	e2_videberta_cross_ce_real_frozenemb_lr1e5	–	argmax	v1	N/A	35.28%	N/A	N/A	N/A	N/A	N/A	22.48%	N/A	N/A	N/A	N/A
–	e2_videberta_cross_ce_real_frozenemb_lr2e5	–	argmax	v1	N/A	35.28%	N/A	N/A	N/A	N/A	N/A	22.48%	N/A	N/A	N/A	N/A
–	e2_videberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_videberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	N/A	35.28%	N/A	N/A	N/A	N/A	N/A	22.48%	N/A	N/A	N/A	N/A
1	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_lv21	jhu-clsp/mmBERT-base	0.076	v2_1	98.14% ±0.37	99.06% ±0.25	87.08% ±2.27	98.19% ±1.61	99.15% ±0.85	99.47% ±0.92	97.36% ±0.26	91.97% ±0.45	60.53% ±3.63	96.28% ±1.15	91.76% ±2.20	66.29% ±4.49
2	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_lv21	vinai/phobert-base-v2	0.033	v2_1	97.03%	98.29%	92.42%	95.82%	98.84%	100.00%	97.17%	91.33%	65.10%	94.95%	86.52%	60.00%

Recall theo từng intent (dev và test_full)

STT	Model_Name	Backbone	Threshold	Data	dev Affirm_Rec	dev Negate_Rec	dev Fallback_Rec	dev Callback_Rec	dev Relation_Rec	dev Repeat_Rec	full Affirm_Rec	full Negate_Rec	full Fallback_Rec	full Callback_Rec	full Relation_Rec	full Repeat_Rec
0	e0_prod_sft_hard_none_qc_v16_m1 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	m1	N/A	N/A	N/A	N/A	N/A	N/A	97.23%	94.72%	69.29%	100.00%	86.96%	100.00%
0	e0_prod_sft_hard_none_qc_v16 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v1	N/A	N/A	N/A	N/A	N/A	N/A	97.19%	94.59%	67.97%	94.59%	85.11%	100.00%
0	e0_prod_sft_hard_none_qc_v16_lv2 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2	N/A	N/A	N/A	N/A	N/A	N/A	97.07%	93.07%	69.01%	98.10%	87.23%	100.00%
0	e0_prod_sft_hard_none_qc_v16_lv21 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2_1	N/A	N/A	N/A	N/A	N/A	N/A	97.19%	94.59%	67.97%	94.59%	85.11%	100.00%
1	e3_ensemble_devtop1_familygeomean_m1	–	0.051	m1	97.05%	98.71%	100.00%	100.00%	98.26%	96.88%	95.67%	85.74%	97.86%	97.98%	88.04%	75.00%
2	e3_ensemble_devtop1_familygeomean_tempdev_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	96.63%	98.61%	100.00%	100.00%	98.26%	96.88%	95.91%	86.09%	97.86%	98.99%	88.04%	75.00%
3	e3_ensemble_devtop1_familygeomean_temphard_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	96.63%	98.61%	100.00%	100.00%	98.26%	96.88%	95.91%	86.09%	97.86%	98.99%	88.04%	75.00%
4	e3_ensemble_devtop1_familymean_m1	–	0.112	m1	96.91%	98.61%	100.00%	100.00%	98.26%	96.88%	96.03%	86.09%	97.14%	98.99%	89.13%	83.33%
5	e3_ensemble_sweep_k2_m1	–	0.110	m1	97.61%	98.50%	100.00%	100.00%	98.26%	96.88%	97.17%	89.08%	94.29%	97.98%	88.04%	83.33%
6	e3_ensemble_sweep_k3_m1	–	0.051	m1	97.05%	98.61%	100.00%	100.00%	98.26%	96.88%	96.45%	87.15%	95.71%	98.99%	86.96%	75.00%
7	e3_ensemble_sweep_k4_m1	–	0.058	m1	97.05%	98.71%	100.00%	100.00%	98.26%	96.88%	96.63%	86.97%	95.71%	98.99%	88.04%	75.00%
8	e3_ensemble_sweep_k5_m1	–	0.111	m1	98.03%	98.82%	100.00%	100.00%	98.26%	96.88%	97.23%	89.08%	96.43%	100.00%	88.04%	83.33%
9	e3_ensemble_sweep_k6_m1	–	0.087	m1	98.03%	98.82%	100.00%	100.00%	98.26%	96.88%	97.05%	88.20%	96.43%	100.00%	88.04%	83.33%
10	e3_ensemble_sweep_k7_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.087	m1	98.03%	98.82%	100.00%	100.00%	98.26%	96.88%	97.05%	88.20%	96.43%	100.00%	88.04%	83.33%
11	e3_ensemble_sweep_k8_m1	–	0.087	m1	98.03%	98.82%	100.00%	100.00%	98.26%	96.88%	97.05%	88.20%	96.43%	100.00%	88.04%	83.33%
12	e3_ensemble_top_geomean_m1	–	0.073	m1	97.19%	98.71%	100.00%	100.00%	98.26%	96.88%	96.03%	87.32%	95.71%	98.99%	90.22%	83.33%
13	e3_ensemble_top_mean_m1	–	0.073	m1	96.21%	98.39%	100.00%	100.00%	97.67%	96.88%	95.07%	85.04%	96.43%	98.99%	90.22%	83.33%
14	e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	96.21% ±0.28	98.21% ±0.50	99.60% ±0.00	100.00% ±0.00	98.16% ±0.44	96.88% ±1.56	94.46% ±1.45	86.56% ±1.48	95.00% ±0.71	97.31% ±2.10	85.14% ±4.12	66.67% ±8.33
15	e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1	intfloat/multilingual-e5-base	0.017	m1	95.65%	98.50%	99.60%	100.00%	97.67%	96.88%	92.78%	86.09%	94.29%	98.99%	81.52%	75.00%
16	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m1	jhu-clsp/mmBERT-base	0.155	m1	96.91%	98.50%	99.60%	100.00%	97.97%	96.88%	96.81%	85.92%	87.14%	94.95%	83.70%	83.33%
17	e2_qwen3_17b_base_genlabel_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.081	m1	96.07%	97.75%	99.60%	98.08%	97.67%	95.31%	94.71%	86.44%	91.43%	98.99%	86.96%	91.67%
18	e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-1.7B-Base	0.105	m1	96.16% ±0.43	97.78% ±0.22	99.60% ±0.00	99.11% ±0.80	98.06% ±0.17	95.83% ±0.90	94.97% ±0.81	86.33% ±1.02	93.81% ±0.41	98.32% ±1.54	89.49% ±1.66	91.67% ±0.00
19	e3_ensemble_top_greedy_m1	–	0.113	m1	96.91%	98.93%	99.60%	100.00%	97.97%	96.88%	95.73%	86.80%	95.00%	98.99%	89.13%	91.67%
20	e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m1	intfloat/multilingual-e5-large-instruct	0.017	m1	96.35% ±0.64	98.43% ±0.27	99.46% ±0.23	99.74% ±0.44	98.16% ±0.17	95.83% ±3.25	95.27% ±0.59	87.21% ±2.52	94.76% ±0.41	97.64% ±0.58	84.78% ±1.88	83.33% ±0.00
21	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1	jhu-clsp/mmBERT-base	0.108	m1	95.79% ±0.74	97.93% ±0.78	99.46% ±0.23	99.11% ±0.96	98.16% ±0.73	95.83% ±0.90	93.08% ±1.39	84.51% ±3.17	93.10% ±5.17	96.30% ±0.58	82.97% ±5.14	83.33% ±14.43
22	e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1	Qwen/Qwen3-4B-Base	0.056	m1	96.02% ±0.08	98.21% ±0.16	99.46% ±0.23	99.74% ±0.22	98.74% ±0.84	96.35% ±0.90	95.91% ±0.48	84.45% ±0.71	95.71% ±0.71	98.65% ±0.58	91.30% ±1.09	80.56% ±4.81
23	e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1	unsloth/gemma-3-1b-pt	0.045	m1	95.22% ±0.40	98.39% ±0.30	99.40% ±0.28	100.00% ±0.00	98.26% ±0.00	96.09% ±1.10	94.95% ±0.26	85.48% ±0.12	95.36% ±0.51	95.96% ±1.43	86.96% ±6.15	75.00% ±0.00
24	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepeval_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1, nhãn m1	vinai/phobert-base-v2	0.033	m1	95.83% ±0.45	98.07% ±0.39	99.33% ±0.23	99.74% ±0.22	97.48% ±0.34	96.88% ±0.00	94.52% ±0.57	84.21% ±0.97	92.38% ±1.49	98.65% ±0.58	86.59% ±2.51	66.67% ±8.33
25	e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1	unsloth/gemma-3-1b-pt	0.045	m1	95.79% ±0.37	98.10% ±0.63	99.33% ±0.61	99.49% ±0.22	98.45% ±0.17	97.40% ±0.90	94.36% ±1.11	83.57% ±1.27	94.29% ±1.24	95.62% ±1.17	88.77% ±3.49	80.56% ±4.81
26	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1	jhu-clsp/mmBERT-base	0.052	m1	96.58% ±0.94	98.10% ±0.22	99.33% ±0.23	99.49% ±0.59	98.06% ±0.17	97.40% ±0.90	94.75% ±0.42	85.39% ±2.59	91.67% ±1.49	95.29% ±2.92	84.06% ±5.36	91.67% ±8.33
27	e2_mmbert_base_cross_ce_synth1x_frozenemb_lr2e5_fbmatched_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	jhu-clsp/mmBERT-base	0.051	m1	96.86% ±0.43	98.28% ±0.11	99.33% ±0.61	99.36% ±0.59	98.06% ±0.61	97.40% ±0.90	95.41% ±0.69	86.56% ±2.15	92.38% ±2.70	95.96% ±1.01	85.14% ±2.74	86.11% ±4.81
28	e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.118	m1	96.35% ±0.28	97.93% ±0.06	99.33% ±0.46	99.87% ±0.22	98.84% ±0.29	96.35% ±0.90	95.49% ±0.96	84.33% ±2.25	93.33% ±3.52	95.96% ±1.01	85.87% ±2.88	83.33% ±8.33
29	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1	vinai/phobert-base-v2	0.008	m1	97.19%	98.39%	99.20%	99.62%	98.55%	89.06%	95.49%	85.74%	93.57%	96.97%	85.87%	16.67%
30	e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1	Qwen/Qwen3-4B-Base	0.162	m1	96.49%	97.85%	99.20%	99.23%	98.55%	96.88%	95.67%	81.51%	96.43%	100.00%	92.39%	100.00%
31	e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1	Qwen/Qwen3-8B-Base	0.170	m1	97.89%	98.61%	99.20%	98.85%	98.26%	96.88%	96.69%	86.80%	92.86%	98.99%	92.39%	91.67%
32	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_m1	vinai/phobert-base-v2	0.073	m1	96.68% ±0.69	98.00% ±0.12	99.20% ±0.40	99.36% ±0.44	98.35% ±0.44	97.40% ±1.80	95.85% ±0.31	82.10% ±0.10	95.00% ±0.71	98.65% ±1.54	89.86% ±1.26	80.56% ±12.73
33	e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	96.63% ±0.56	98.21% ±0.12	99.20% ±0.70	99.49% ±0.59	98.26% ±0.29	97.40% ±0.90	95.49% ±0.52	85.45% ±0.57	93.57% ±2.58	96.97% ±1.01	86.59% ±6.64	75.00% ±0.00
34	e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-base	0.031	m1	95.27% ±0.58	97.60% ±0.41	99.20% ±0.40	99.36% ±0.59	98.74% ±0.34	95.83% ±0.90	94.40% ±0.84	85.97% ±2.12	95.48% ±1.65	98.32% ±0.58	84.78% ±2.88	63.89% ±4.81
35	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.061	m1	96.02% ±1.14	98.28% ±0.21	99.20% ±0.80	99.87% ±0.22	98.06% ±0.44	96.88% ±0.00	94.65% ±0.63	84.92% ±2.95	94.05% ±1.09	96.63% ±1.17	83.70% ±4.98	77.78% ±4.81
36	e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1	Qwen/Qwen3-4B-Base	0.131	m1	96.07% ±1.34	98.14% ±0.38	99.20% ±0.70	99.74% ±0.22	98.35% ±0.44	96.88% ±0.00	96.53% ±1.04	85.62% ±1.32	94.76% ±2.51	98.99% ±1.01	89.86% ±1.66	86.11% ±4.81
37	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.043	m1	95.79% ±0.85	97.71% ±0.53	99.06% ±0.61	99.87% ±0.22	97.48% ±0.44	93.75% ±1.56	94.95% ±0.43	84.62% ±0.37	92.38% ±2.06	98.32% ±1.17	87.32% ±4.39	72.22% ±9.62
38	e2_gemma3_1b_pt_cross_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.053	m1	95.93% ±0.61	98.07% ±0.21	99.06% ±0.23	100.00% ±0.00	98.06% ±0.17	95.83% ±1.80	95.25% ±0.48	85.45% ±1.50	94.29% ±0.00	95.96% ±1.01	85.14% ±2.51	75.00% ±0.00
39	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1	vinai/phobert-base-v2	0.037	m1	96.11% ±0.66	98.28% ±0.19	98.93% ±0.46	99.62% ±0.38	97.58% ±0.44	96.35% ±2.39	95.35% ±1.04	84.27% ±1.74	93.33% ±1.65	97.31% ±0.58	88.04% ±1.88	83.33% ±8.33
40	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m1	vinai/phobert-base-v2	0.085	m1	96.77% ±1.34	98.21% ±0.12	98.93% ±0.93	99.62% ±0.00	97.97% ±0.77	97.40% ±1.80	96.21% ±0.43	84.62% ±1.74	92.14% ±0.71	98.99% ±0.00	88.04% ±1.88	80.56% ±12.73
41	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_m1	vinai/phobert-base-v2	0.028	m1	96.11% ±0.77	97.96% ±0.57	98.93% ±0.46	99.74% ±0.44	97.77% ±0.44	97.40% ±0.90	95.17% ±0.88	83.33% ±1.87	93.57% ±1.24	97.64% ±0.58	85.87% ±1.09	80.56% ±20.97
42	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1	jhu-clsp/mmBERT-base	0.082	m1	96.58% ±0.43	98.00% ±0.96	98.93% ±0.46	99.49% ±0.22	98.26% ±0.00	96.35% ±0.90	95.25% ±0.65	84.57% ±3.32	92.38% ±3.60	95.96% ±1.01	82.97% ±3.32	80.56% ±4.81
43	e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.057	m1	96.07% ±0.78	98.07% ±0.47	98.93% ±0.84	99.11% ±0.44	98.35% ±0.44	96.88% ±0.00	94.56% ±0.39	84.51% ±1.38	93.10% ±2.18	95.62% ±2.54	84.06% ±3.32	88.89% ±4.81
44	e2_qwen3_17b_base_cross_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.101	m1	95.41% ±1.06	98.28% ±0.11	98.93% ±0.61	99.49% ±0.88	98.55% ±0.50	95.31% ±0.00	94.83% ±0.68	85.04% ±0.61	91.90% ±2.97	97.98% ±2.02	89.13% ±2.17	80.56% ±9.62
45	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m1	vinai/phobert-base-v2	0.052	m1	95.97% ±1.02	98.14% ±0.54	98.80% ±0.40	99.74% ±0.22	97.48% ±0.44	94.27% ±3.25	95.07% ±0.78	84.74% ±3.40	93.81% ±1.09	97.98% ±1.01	88.04% ±2.17	75.00% ±14.43
46	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1	vinai/phobert-base-v2	0.077	m1	95.93% ±0.84	97.64% ±0.11	98.80% ±0.70	99.74% ±0.22	98.45% ±0.34	95.83% ±0.90	95.77% ±0.76	84.21% ±3.00	94.52% ±0.41	98.65% ±0.58	90.22% ±2.17	72.22% ±9.62
47	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	vinai/phobert-base-v2	0.058	m1	95.60% ±0.57	97.96% ±0.28	98.80% ±0.40	99.62% ±0.38	97.19% ±0.17	95.31% ±3.12	95.07% ±0.12	84.62% ±0.73	92.86% ±0.71	97.31% ±2.10	87.68% ±3.32	80.56% ±12.73
48	e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1	unsloth/gemma-3-1b-pt	0.056	m1	96.11% ±0.21	98.18% ±0.21	98.80% ±0.70	99.74% ±0.44	98.26% ±0.29	97.40% ±0.90	95.39% ±0.46	83.74% ±1.71	93.57% ±0.71	96.63% ±1.17	84.78% ±3.92	80.56% ±4.81
49	e2_gemma3_270m_cross_ce_real_lora_m1gemma-3-270m · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn m1	unsloth/gemma-3-270m	0.098	m1	95.97% ±0.71	97.71% ±0.34	98.80% ±0.00	99.49% ±0.88	97.67% ±0.77	97.92% ±0.90	94.71% ±0.53	82.10% ±1.59	91.67% ±1.65	95.62% ±0.58	80.80% ±5.99	83.33% ±8.33
50	e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-large-instruct	0.029	m1	96.63% ±0.88	98.32% ±0.27	98.80% ±1.06	99.74% ±0.22	98.16% ±0.34	96.35% ±0.90	95.55% ±0.51	86.62% ±0.81	93.33% ±1.49	98.65% ±0.58	84.06% ±2.74	72.22% ±9.62
51	e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.034	m1	95.51% ±0.28	98.07% ±0.21	98.80% ±0.80	99.87% ±0.22	98.16% ±0.17	95.83% ±0.90	93.40% ±1.62	85.97% ±1.58	94.29% ±1.43	97.31% ±1.17	84.42% ±1.66	94.44% ±4.81
52	e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.051	m1	95.69% ±0.21	98.18% ±0.19	98.80% ±0.40	99.49% ±0.59	98.06% ±0.34	95.31% ±0.00	94.14% ±0.21	86.44% ±0.91	93.57% ±0.71	93.94% ±5.34	82.97% ±2.51	88.89% ±12.73
53	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1	jhu-clsp/mmBERT-base	0.040	m1	96.40% ±0.66	97.78% ±0.43	98.80% ±1.39	99.74% ±0.22	98.35% ±0.44	95.83% ±1.80	94.34% ±0.89	84.62% ±2.00	94.52% ±0.82	95.96% ±2.67	82.25% ±5.47	72.22% ±4.81
54	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1	jhu-clsp/mmBERT-base	0.057	m1	95.69% ±0.21	97.85% ±0.19	98.80% ±1.06	98.60% ±0.96	98.26% ±0.29	96.88% ±0.00	94.81% ±1.38	85.15% ±0.37	94.05% ±1.09	96.63% ±1.54	84.42% ±2.26	83.33% ±0.00
55	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.070	m1	96.63%	98.71%	98.80%	99.62%	97.97%	96.88%	95.55%	86.80%	92.14%	96.97%	83.70%	91.67%
56	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.047	m1	96.91% ±0.56	97.85% ±0.43	98.80% ±1.45	99.87% ±0.22	98.45% ±0.17	96.35% ±1.80	94.87% ±0.72	85.39% ±1.15	91.90% ±0.82	96.30% ±2.54	84.42% ±3.14	75.00% ±8.33
57	e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1	Qwen/Qwen3-1.7B-Base	0.095	m1	97.00% ±0.58	97.96% ±0.28	98.80% ±0.80	98.60% ±1.23	97.77% ±0.17	94.79% ±0.90	95.91% ±0.28	86.27% ±2.91	91.90% ±1.80	96.97% ±1.75	88.77% ±2.26	80.56% ±12.73
58	e2_xlmr_base_khonly_ce_real_frozenemb_m1XLM-R base · chỉ câu KH · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1	FacebookAI/xlm-roberta-base	0.012	m1	94.52%	97.42%	98.80%	100.00%	98.26%	93.75%	94.04%	83.10%	95.00%	98.99%	88.04%	66.67%
59	e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.042	m1	96.25% ±0.29	97.71% ±0.22	98.66% ±0.61	99.49% ±0.22	97.58% ±0.34	96.35% ±1.80	94.71% ±0.75	84.80% ±1.06	94.29% ±1.89	95.96% ±1.75	82.61% ±1.88	80.56% ±9.62
60	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m1	vinai/phobert-base-v2	0.024	m1	95.65% ±0.28	98.00% ±0.27	98.66% ±0.23	99.74% ±0.22	97.58% ±0.17	94.79% ±2.39	94.83% ±0.16	83.33% ±1.25	94.52% ±1.09	99.66% ±0.58	86.59% ±6.37	72.22% ±9.62
61	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m1	vinai/phobert-base-v2	0.031	m1	96.72% ±0.21	97.75% ±0.60	98.66% ±0.61	99.62% ±0.38	97.67% ±0.29	94.27% ±4.51	95.75% ±0.58	82.92% ±2.79	92.62% ±2.06	97.98% ±1.75	85.51% ±6.28	75.00% ±14.43
62	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m1	vinai/phobert-base-v2	0.031	m1	95.83% ±0.21	97.82% ±0.38	98.66% ±0.46	99.74% ±0.44	97.29% ±0.34	95.31% ±1.56	95.33% ±0.94	84.74% ±0.57	94.52% ±0.41	97.64% ±1.17	84.78% ±3.26	75.00% ±8.33
63	e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1	unsloth/gemma-3-1b-pt	0.096	m1	96.16% ±0.58	98.10% ±0.12	98.66% ±0.46	99.62% ±0.66	98.16% ±0.44	97.92% ±0.90	95.27% ±0.91	84.86% ±1.96	91.43% ±0.00	96.63% ±1.17	87.68% ±3.32	55.56% ±26.79
64	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1	jhu-clsp/mmBERT-base	0.061	m1	95.55% ±1.05	97.89% ±0.33	98.66% ±0.84	99.36% ±0.22	98.35% ±0.34	95.83% ±1.80	93.10% ±0.79	85.09% ±1.66	94.29% ±1.89	96.30% ±3.09	84.78% ±3.26	80.56% ±20.97
65	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.017	m1	97.10% ±0.29	98.07% ±0.21	98.66% ±0.61	99.87% ±0.22	98.35% ±0.34	92.71% ±3.25	95.57% ±0.17	85.62% ±1.00	93.10% ±1.65	97.64% ±0.58	86.59% ±2.26	58.33% ±0.00
66	e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1	vinai/phobert-base-v2	0.026	m1	97.19% ±0.14	98.00% ±0.16	98.66% ±0.23	100.00% ±0.00	98.06% ±0.17	95.31% ±1.56	95.95% ±0.57	86.21% ±1.42	91.43% ±1.24	98.32% ±0.58	86.23% ±0.63	66.67% ±8.33
67	e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.065	m1	95.51% ±0.56	98.03% ±0.16	98.53% ±0.23	99.74% ±0.22	97.97% ±0.50	95.83% ±0.90	95.35% ±0.60	84.10% ±1.50	93.10% ±1.09	97.31% ±1.54	86.59% ±0.63	77.78% ±4.81
68	e2_xlmr_base_cross_logitadjust_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.018	m1	95.51% ±0.51	97.46% ±0.33	98.53% ±0.23	99.62% ±0.00	98.35% ±0.34	95.83% ±1.80	94.20% ±0.51	81.34% ±1.53	93.81% ±1.49	98.99% ±0.00	87.68% ±2.51	75.00% ±0.00
69	e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-large-instruct	0.026	m1	96.40% ±0.89	98.64% ±0.06	98.53% ±0.84	100.00% ±0.00	98.06% ±0.17	96.35% ±0.90	95.37% ±0.93	88.62% ±0.97	92.62% ±1.49	97.31% ±1.54	86.23% ±1.26	83.33% ±0.00
70	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1	jhu-clsp/mmBERT-base	0.072	m1	97.14% ±0.69	97.96% ±0.39	98.53% ±1.16	99.49% ±0.59	98.26% ±0.29	96.88% ±1.56	95.63% ±0.89	85.97% ±2.06	90.00% ±2.47	95.96% ±2.67	81.88% ±6.55	75.00% ±8.33
71	e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1	jhu-clsp/mmBERT-base	0.082	m1	95.79% ±0.37	98.21% ±0.33	98.39% ±0.80	99.11% ±0.80	97.87% ±0.17	96.88% ±0.00	95.25% ±0.64	85.92% ±1.61	93.33% ±1.80	94.61% ±2.54	81.52% ±7.61	77.78% ±9.62
72	e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.067	m1	95.51%	97.85%	98.39%	99.62%	98.26%	95.31%	93.74%	83.10%	92.86%	97.98%	85.87%	83.33%
73	e2_visobert_cross_ce_real_frozenemb_lr4e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	uitnlp/visobert	0.037	m1	95.51%	98.18%	98.39%	99.23%	98.55%	95.31%	91.70%	82.04%	93.57%	93.94%	84.78%	75.00%
74	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1	FacebookAI/xlm-roberta-base	0.027	m1	95.37%	97.85%	98.39%	100.00%	98.26%	93.75%	93.20%	86.09%	88.57%	98.99%	85.87%	75.00%
75	e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m1mmBERT base · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.057	m1	95.55% ±0.53	97.89% ±0.16	98.39% ±1.06	99.87% ±0.22	98.06% ±0.44	96.88% ±2.71	94.89% ±0.49	85.74% ±2.83	94.05% ±1.80	96.30% ±0.58	83.70% ±3.26	66.67% ±8.33
76	e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.206	m1	96.86% ±1.00	98.10% ±0.22	98.39% ±0.40	99.74% ±0.22	97.67% ±0.58	97.40% ±0.90	96.41% ±1.22	86.38% ±3.35	89.29% ±5.00	98.32% ±2.10	88.41% ±0.63	86.11% ±9.62
77	e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.034	m1	95.83% ±0.21	97.82% ±0.34	98.39% ±0.70	100.00% ±0.00	97.48% ±0.67	94.79% ±2.39	95.85% ±0.26	85.33% ±0.83	91.19% ±1.09	99.33% ±0.58	88.77% ±1.26	75.00% ±14.43
78	e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.052	m1	96.40% ±0.63	97.85% ±0.43	98.26% ±0.23	100.00% ±0.00	97.29% ±0.17	96.88% ±0.00	96.23% ±1.01	84.33% ±1.86	92.14% ±3.98	98.65% ±0.58	83.70% ±4.98	66.67% ±0.00
79	e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.033	m1	96.07% ±1.10	97.75% ±0.28	98.26% ±0.23	99.87% ±0.22	97.38% ±0.00	94.79% ±0.90	96.07% ±0.70	84.10% ±2.20	93.10% ±2.30	97.98% ±0.00	82.97% ±5.14	66.67% ±8.33
80	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1	jhu-clsp/mmBERT-base	0.144	m1	96.40% ±0.77	98.07% ±0.54	98.26% ±1.41	98.72% ±0.80	98.26% ±0.29	95.83% ±1.80	94.65% ±0.48	86.03% ±0.97	92.14% ±0.71	93.60% ±2.33	85.87% ±1.88	75.00% ±8.33
81	e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.082	m1	96.72% ±0.86	98.00% ±0.27	98.26% ±1.29	99.87% ±0.22	98.45% ±0.34	96.88% ±0.00	94.69% ±1.48	85.15% ±1.78	91.67% ±4.31	98.99% ±1.01	84.06% ±2.74	97.22% ±4.81
82	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1	jhu-clsp/mmBERT-base	0.108	m1	96.54% ±0.57	98.03% ±0.27	98.26% ±0.61	98.98% ±1.11	98.26% ±0.29	95.31% ±0.00	95.93% ±0.82	86.56% ±1.03	92.38% ±2.18	96.30% ±1.54	82.97% ±3.49	80.56% ±4.81
83	e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-0.6B-Base	0.075	m1	95.93% ±0.74	98.03% ±0.16	98.13% ±1.29	100.00% ±0.00	97.00% ±0.34	98.44% ±1.56	94.62% ±1.19	83.80% ±1.79	94.05% ±1.49	98.32% ±0.58	89.86% ±2.26	86.11% ±4.81
84	e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1	Qwen/Qwen3-0.6B-Base	0.077	m1	96.02% ±0.29	97.35% ±0.61	97.99% ±1.61	97.57% ±0.22	97.97% ±0.58	95.31% ±0.00	92.94% ±0.71	83.57% ±3.33	90.95% ±3.60	96.30% ±1.17	87.68% ±2.74	77.78% ±9.62
85	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_m1	vinai/phobert-base-v2	0.093	m1	95.41% ±1.27	97.78% ±0.75	97.99% ±0.40	99.74% ±0.22	97.77% ±0.17	97.40% ±0.90	95.01% ±1.41	86.91% ±3.42	93.10% ±0.41	99.66% ±0.58	86.23% ±1.26	88.89% ±12.73
86	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m1	vinai/phobert-base-v2	0.104	m1	96.63% ±0.51	97.78% ±0.27	97.99% ±1.45	99.62% ±0.00	98.06% ±0.61	95.83% ±0.90	95.87% ±1.09	85.56% ±2.00	92.14% ±2.86	98.99% ±1.01	83.33% ±3.32	77.78% ±9.62
87	e2_xlmr_base_cross_ce_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.044	m1	95.79% ±0.51	97.60% ±0.45	97.99% ±0.40	100.00% ±0.00	98.55% ±0.29	92.71% ±1.80	94.52% ±0.96	83.04% ±3.91	90.95% ±3.30	97.31% ±1.17	88.04% ±1.09	69.44% ±12.73
88	e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.019	m1	95.83% ±0.29	97.64% ±0.21	97.99% ±0.70	99.87% ±0.22	98.35% ±0.17	97.40% ±1.80	94.58% ±0.90	82.98% ±1.34	93.10% ±2.97	98.65% ±0.58	86.59% ±2.74	83.33% ±8.33
89	e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.049	m1	95.32% ±0.69	97.82% ±0.27	97.99% ±1.75	99.74% ±0.22	97.58% ±0.44	94.79% ±2.39	95.13% ±0.26	83.80% ±1.79	93.10% ±2.06	97.98% ±1.01	86.59% ±1.66	77.78% ±9.62
90	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.159	m1	95.60% ±0.35	97.42% ±0.47	97.72% ±1.16	99.74% ±0.22	97.48% ±0.34	93.23% ±0.90	95.69% ±0.72	84.51% ±2.71	91.67% ±1.49	98.99% ±1.01	84.42% ±5.58	61.11% ±31.55
91	e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-4B	0.061	m1	94.24% ±0.60	96.62% ±2.35	97.39% ±0.85	96.93% ±1.63	97.67% ±0.00	97.66% ±1.10	94.49% ±0.47	81.43% ±2.12	90.71% ±3.03	88.89% ±8.57	79.35% ±4.61	83.33% ±11.79
92	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.025	m1	97.19%	97.10%	97.19%	100.00%	97.67%	89.06%	96.15%	84.51%	90.00%	97.98%	82.61%	16.67%
93	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.152	m1	95.32% ±0.35	96.85% ±0.66	97.05% ±1.23	99.23% ±0.66	98.16% ±0.61	97.40% ±0.90	94.77% ±0.48	83.98% ±3.07	85.24% ±0.41	93.94% ±0.00	80.07% ±4.12	86.11% ±4.81
94	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m1	vinai/phobert-base-v2	0.035	m1	95.69% ±0.08	98.10% ±0.27	96.92% ±1.85	99.62% ±0.38	97.77% ±0.67	94.79% ±0.90	95.07% ±1.03	84.74% ±2.22	94.05% ±1.49	98.32% ±1.54	87.68% ±2.74	77.78% ±4.81
95	e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-base	0.052	m1	94.10%	97.42%	96.79%	100.00%	98.26%	93.75%	93.98%	85.39%	90.71%	96.97%	82.61%	25.00%
96	e2_visobert_cross_ce_real_frozenemb_lr2e5_m1	uitnlp/visobert	0.027	m1	95.37%	97.42%	96.39%	99.62%	98.26%	93.75%	92.78%	80.28%	90.71%	94.95%	86.96%	58.33%
97	e2_visobert_cross_ce_real_frozenemb_lr1e5_m1	uitnlp/visobert	0.186	m1	93.54%	96.78%	95.18%	99.62%	97.97%	93.75%	92.84%	82.39%	87.14%	94.95%	77.17%	66.67%
98	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1	FacebookAI/xlm-roberta-base	0.161	m1	95.22%	97.00%	95.18%	100.00%	98.55%	90.62%	94.22%	83.98%	90.00%	97.98%	81.52%	58.33%
99	e1_tfidf_hybrid_qwen3emb_lr_m1TF-IDF câu KH (char 1-5, word 1-2-gram, giữ từ 1 ký tự, thêm kênh char bỏ dấu) + câu bot (char 2-4) + tương tác loại câu bot × từ câu KH · vector đóng băng Qwen3-Embedding-0.6B (cặp bot + KH, pooling last_token), trọng số dense 1 · LogReg · train thật, nhãn m1	Qwen/Qwen3-Embedding-0.6B	0.149	m1	94.10%	96.57%	94.38%	98.85%	97.38%	98.44%	94.77%	79.75%	85.71%	89.90%	79.35%	66.67%
100	e1_tfidf_svmcal_ce_synth1x_fbmatched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	TF-IDF	0.207	m1	93.82%	96.78%	92.77%	98.85%	97.38%	95.31%	95.49%	84.68%	82.86%	94.95%	79.35%	66.67%
101	e1_tfidf_svmcal_ce_synth1x_nofb_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent omi_intent_repeat, callback, relationship, negate, affirm), nhãn m1	TF-IDF	0.164	m1	93.40%	96.67%	92.77%	99.62%	97.38%	95.31%	95.43%	84.33%	82.14%	94.95%	79.35%	66.67%
102	e1_tfidf_svmcal_ce_synth1x_noqparticle_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm q_particle), nhãn m1	TF-IDF	0.280	m1	93.40%	96.67%	92.37%	98.85%	97.38%	95.31%	93.74%	82.75%	85.71%	94.95%	78.26%	83.33%
103	e1_tfidf_svmcal_ce_real_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn m1	TF-IDF	0.178	m1	93.54%	97.00%	91.97%	99.62%	97.09%	95.31%	94.89%	85.21%	85.00%	94.95%	79.35%	66.67%
104	e1_tfidf_svmcal_ce_synth1x_nofiller_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm filler_only), nhãn m1	TF-IDF	0.267	m1	93.82%	96.46%	91.97%	99.23%	97.38%	95.31%	94.04%	81.87%	88.57%	94.95%	77.17%	75.00%
105	e1_tfidf_svmcal_ce_synth1x_noofftopic_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm prefix_offtopic), nhãn m1	TF-IDF	0.298	m1	93.40%	96.67%	91.97%	99.23%	97.38%	95.31%	93.86%	82.92%	85.71%	94.95%	79.35%	83.33%
106	e2_dual_qwen3emb06b_concat_ce_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.227	m1	95.22% ±0.84	95.82% ±0.43	91.83% ±0.23	98.72% ±0.88	95.74% ±0.61	96.35% ±0.90	93.72% ±1.27	78.29% ±0.44	86.90% ±1.09	91.92% ±3.03	69.20% ±6.64	66.67% ±8.33
107	e1_tfidf_svmcal_ce_synth1x_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1, nhãn m1	TF-IDF	0.282	m1	93.54%	96.89%	91.57%	98.85%	97.38%	95.31%	93.80%	82.22%	85.00%	94.95%	78.26%	83.33%
108	e1_tfidf_svmcal_ce_synth1x_matched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ intent của train thật), nhãn m1	TF-IDF	0.190	m1	93.26%	97.00%	91.57%	99.23%	97.09%	95.31%	94.95%	85.21%	84.29%	94.95%	78.26%	66.67%
109	e1_tfidf_svmcal_ce_synth1x_nomoney_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm money), nhãn m1	TF-IDF	0.310	m1	93.96%	96.67%	91.57%	98.85%	97.38%	95.31%	93.92%	83.10%	86.43%	94.95%	80.43%	83.33%
110	e1_tfidf_svmcal_ce_synth1x_fbonly_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent nlu_fallback), nhãn m1	TF-IDF	0.356	m1	93.96%	97.21%	91.16%	99.62%	97.09%	95.31%	92.60%	83.10%	86.43%	94.95%	79.35%	83.33%
–	e2_bamibert_cross_ce_real_frozenemb_lr1e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_bamibert_cross_ce_real_frozenemb_lr2e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_bamibert_cross_ce_real_frozenemb_lr4e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e3_fbverifier_devtop1_familygeomean_fbf1_m1	–	hard	m1	98.31%	98.93%	95.98%	100.00%	98.26%	98.44%	98.01%	91.55%	86.43%	100.00%	89.13%	91.67%
–	e3_fbverifier_devtop1_familygeomean_macrof1_m1	–	hard	m1	98.31%	98.93%	95.98%	100.00%	98.26%	98.44%	98.01%	91.55%	86.43%	100.00%	89.13%	91.67%
1	e3_ensemble_mean_all	–	0.065	v1	96.37%	98.06%	98.85%	99.62%	99.43%	92.31%	96.29%	85.14%	90.20%	90.99%	86.17%	83.33%
2	e3_ensemble_greedy_all	–	0.070	v1	96.02%	98.16%	98.47%	98.87%	99.43%	95.38%	96.23%	85.14%	89.54%	90.99%	85.11%	83.33%
3	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5	–	0.023	v1	97.42%	97.65%	98.09%	99.62%	97.99%	93.85%	96.77%	86.15%	86.27%	90.99%	81.91%	83.33%
4	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1	vinai/phobert-base-v2	0.011	v1	96.84%	98.06%	97.71%	99.25%	97.71%	93.85%	97.01%	82.77%	86.27%	90.99%	87.23%	66.67%
5	e2_xlmr_base_cross_wcesqrt_real_frozenemb	FacebookAI/xlm-roberta-base	0.021	v1	95.39% ±0.24	97.38% ±0.31	97.71% ±0.38	99.75% ±0.43	99.43% ±0.00	93.33% ±0.89	96.01% ±0.03	85.14% ±1.61	87.58% ±1.31	91.59% ±0.52	86.88% ±1.63	80.56% ±4.81
6	e3_prior_em_mmbert_bottype	–	0.082	v1	96.49%	97.85%	97.71%	98.50%	99.14%	93.85%	96.59%	62.50%	85.62%	86.49%	80.85%	66.67%
7	e3_prior_em_mmbert_global	–	0.082	v1	96.49%	97.85%	97.71%	98.50%	99.14%	93.85%	97.49%	82.26%	88.24%	89.19%	79.79%	83.33%
8	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5	jhu-clsp/mmBERT-base	0.061	v1	96.84% ±0.23	97.65% ±0.57	97.58% ±0.58	98.62% ±0.22	98.95% ±0.44	92.82% ±1.78	96.55% ±0.18	84.57% ±2.02	88.24% ±0.65	89.19% ±0.90	87.23% ±3.84	94.44% ±4.81
9	e2_xlmr_base_cross_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	95.98% ±0.88	97.99% ±0.72	97.46% ±0.22	99.87% ±0.22	98.76% ±0.17	92.31% ±0.00	96.45% ±0.09	86.43% ±1.28	86.27% ±2.61	90.99% ±0.00	85.46% ±2.21	83.33% ±0.00
10	e2_xlmr_base_cross_logitadjust_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	96.02% ±0.65	97.55% ±0.18	97.46% ±0.58	100.00% ±0.00	99.24% ±0.60	94.36% ±1.78	96.07% ±0.30	84.97% ±0.77	87.80% ±2.30	91.89% ±0.90	86.88% ±1.63	86.11% ±4.81
11	e2_qwen3_06b_base_cross_ce_real_lora	Qwen/Qwen3-0.6B-Base	0.098	v1	96.02%	97.85%	97.33%	98.87%	97.99%	95.38%	95.81%	86.15%	87.58%	90.09%	86.17%	83.33%
12	e3_ensemble_stack_all	–	0.088	v1	95.32%	98.47%	97.33%	98.12%	99.14%	90.77%	95.63%	87.67%	88.24%	88.29%	84.04%	83.33%
13	e2_qwen3_17b_base_cross_ce_real_lora	Qwen/Qwen3-1.7B-Base	0.083	v1	95.90% ±0.20	97.62% ±0.33	97.20% ±0.58	97.62% ±1.85	98.85% ±0.29	92.82% ±1.78	96.21% ±0.19	84.80% ±1.05	86.93% ±1.31	89.79% ±0.52	86.88% ±2.21	94.44% ±9.62
14	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5	vinai/phobert-base-v2	0.022	v1	96.29% ±0.07	97.85% ±0.27	97.07% ±0.96	100.00% ±0.00	97.90% ±0.60	90.26% ±3.87	96.73% ±0.12	85.87% ±0.59	87.15% ±1.00	90.69% ±0.52	86.17% ±1.84	69.44% ±9.62
15	e2_me5_base_cross_ce_real_frozenemb_lr1e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	intfloat/multilingual-e5-base	0.088	v1	95.43%	97.55%	96.95%	100.00%	98.85%	92.31%	95.63%	85.64%	89.54%	89.19%	85.11%	75.00%
16	e2_me5_base_cross_ce_real_frozenemb_lr4e5	–	0.028	v1	95.32%	98.06%	96.95%	98.87%	99.43%	92.31%	96.53%	85.64%	88.24%	90.99%	85.11%	83.33%
17	e2_xlmr_base_khonly_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.076	v1	95.28% ±0.69	97.31% ±0.33	96.69% ±0.58	99.12% ±0.43	98.76% ±0.33	93.33% ±0.89	96.33% ±0.45	85.81% ±1.61	86.06% ±2.30	90.69% ±1.04	86.52% ±2.46	80.56% ±9.62
18	e2_me5_base_cross_ce_real_frozenemb_lr2e5	intfloat/multilingual-e5-base	0.056	v1	95.24% ±0.47	98.20% ±0.06	96.56% ±0.76	99.75% ±0.43	99.43% ±0.29	93.85% ±3.08	95.97% ±0.31	87.33% ±0.34	87.36% ±1.36	91.29% ±0.52	85.82% ±1.23	77.78% ±4.81
19	e2_gemma3_270m_cross_ce_real_lora	unsloth/gemma-3-270m	0.116	v1	95.47% ±0.36	97.82% ±0.43	96.44% ±0.88	98.37% ±0.87	98.28% ±0.57	95.38% ±0.00	95.87% ±0.84	85.36% ±0.26	88.24% ±2.36	88.89% ±0.52	82.62% ±2.21	86.11% ±4.81
20	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen	Qwen/Qwen3-Embedding-0.6B	0.114	v1	95.43% ±0.12	96.83% ±0.18	96.06% ±1.23	98.87% ±0.38	98.28% ±0.50	94.36% ±0.89	95.83% ±0.14	84.29% ±0.61	84.97% ±2.85	89.19% ±1.80	86.88% ±0.61	91.67% ±0.00
21	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5	–	0.075	v1	97.19%	97.85%	95.80%	98.87%	98.57%	92.31%	96.65%	85.81%	87.58%	89.19%	86.17%	83.33%
22	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5	–	0.026	v1	96.37%	96.83%	95.80%	100.00%	97.71%	81.54%	96.53%	83.28%	88.89%	90.09%	85.11%	16.67%
23	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	FacebookAI/xlm-roberta-base	0.123	v1	95.67%	97.14%	95.80%	100.00%	97.71%	81.54%	96.35%	87.33%	84.31%	92.79%	80.85%	16.67%
24	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5	jhu-clsp/mmBERT-base	0.120	v1	95.84% ±0.25	97.91% ±0.36	95.61% ±0.27	98.12% ±0.00	99.00% ±1.01	95.38% ±0.00	96.17% ±0.17	86.66% ±0.72	85.29% ±1.39	90.09% ±0.00	88.30% ±4.51	83.33% ±0.00
25	e2_visobert_cross_ce_real_frozenemb_lr2e5	–	0.093	v1	95.32%	95.51%	95.04%	97.74%	99.43%	92.31%	95.81%	86.49%	83.66%	88.29%	84.04%	83.33%
26	e2_visobert_cross_ce_real_frozenemb_lr4e5	–	0.066	v1	95.32%	97.14%	94.66%	97.37%	99.71%	92.31%	95.45%	85.98%	84.97%	89.19%	85.11%	91.67%
27	e1_tfidf_svmcal_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn v1	TF-IDF	0.123	v1	94.50%	95.20%	94.27%	98.87%	96.56%	95.38%	96.05%	83.28%	88.89%	89.19%	80.85%	83.33%
28	e1_tfidf_logreg_balanced_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg cân bằng lớp · train thật, nhãn v1	TF-IDF	0.256	v1	94.38%	94.89%	93.89%	98.87%	98.28%	96.92%	95.39%	78.04%	86.27%	88.29%	88.30%	91.67%
29	e1_tfidf_logreg_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật, nhãn v1	TF-IDF	0.125	v1	94.61%	95.20%	93.89%	98.87%	97.71%	93.85%	95.51%	81.25%	88.24%	87.39%	79.79%	83.33%
30	e2_visobert_cross_ce_real_frozenemb_lr1e5	–	0.232	v1	94.38%	94.69%	92.75%	98.87%	98.28%	90.77%	96.17%	85.30%	78.43%	88.29%	81.91%	83.33%
31	e1_tfidf_logreg_ce_synth1x_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật + synth ×1, nhãn v1	TF-IDF	0.234	v1	94.85%	94.99%	92.37%	98.87%	96.56%	92.31%	96.05%	82.26%	84.97%	88.29%	79.79%	83.33%
–	e2_mdeberta_cross_ce_real_frozenemb_lr1e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%
–	e2_videberta_cross_ce_real_frozenemb_lr1e5	–	argmax	v1	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%
–	e2_videberta_cross_ce_real_frozenemb_lr2e5	–	argmax	v1	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%
–	e2_videberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_videberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%	0.00%	100.00%	0.00%	0.00%	0.00%	0.00%
1	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_lv21	jhu-clsp/mmBERT-base	0.076	v2_1	97.10% ±0.72	97.58% ±0.57	96.83% ±0.69	97.91% ±0.60	98.95% ±0.60	94.87% ±0.89	96.29% ±0.49	88.40% ±0.70	82.35% ±1.31	85.29% ±0.52	86.52% ±3.42	86.11% ±12.73
2	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_lv21	vinai/phobert-base-v2	0.033	v2_1	96.24%	97.61%	96.83%	98.82%	97.71%	96.92%	96.59%	89.02%	81.70%	84.68%	81.91%	75.00%

F1 theo từng intent (dev và test_full)

STT	Model_Name	Backbone	Threshold	Data	dev Affirm_F1	dev Negate_F1	dev Fallback_F1	dev Callback_F1	dev Relation_F1	dev Repeat_F1	full Affirm_F1	full Negate_F1	full Fallback_F1	full Callback_F1	full Relation_F1	full Repeat_F1
0	e0_prod_sft_hard_none_qc_v16_m1 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	m1	N/A	N/A	N/A	N/A	N/A	N/A	97.53%	91.81%	76.98%	98.02%	90.40%	82.76%
0	e0_prod_sft_hard_none_qc_v16 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v1	N/A	N/A	N/A	N/A	N/A	N/A	97.31%	91.58%	76.19%	95.02%	89.39%	75.00%
0	e0_prod_sft_hard_none_qc_v16_lv2 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2	N/A	N/A	N/A	N/A	N/A	N/A	97.22%	91.19%	74.81%	95.81%	91.62%	82.35%
0	e0_prod_sft_hard_none_qc_v16_lv21 tham chiếu	Qwen-2B-Chat (SFT generative)	N/A (CausalLM)	v2_1	N/A	N/A	N/A	N/A	N/A	N/A	97.31%	91.58%	76.19%	95.02%	89.39%	75.00%
1	e3_ensemble_devtop1_familygeomean_m1	–	0.051	m1	97.81%	99.08%	95.04%	100.00%	99.12%	98.41%	97.31%	90.19%	69.37%	96.52%	90.50%	78.26%
2	e3_ensemble_devtop1_familygeomean_tempdev_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	97.45%	99.03%	94.68%	100.00%	99.12%	98.41%	97.43%	90.39%	70.62%	97.03%	90.50%	78.26%
3	e3_ensemble_devtop1_familygeomean_temphard_m1ensemble trung bình nhân theo nhóm trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.051	m1	97.45%	99.03%	94.68%	100.00%	99.12%	98.41%	97.43%	90.39%	70.62%	97.03%	90.50%	78.26%
4	e3_ensemble_devtop1_familymean_m1	–	0.112	m1	97.60%	98.92%	95.40%	100.00%	99.12%	98.41%	97.38%	90.47%	71.20%	97.03%	91.62%	83.33%
5	e3_ensemble_sweep_k2_m1	–	0.110	m1	97.82%	98.98%	96.14%	100.00%	99.12%	98.41%	97.70%	92.00%	79.04%	94.63%	92.05%	80.00%
6	e3_ensemble_sweep_k3_m1	–	0.051	m1	97.81%	99.14%	94.50%	100.00%	99.12%	98.41%	97.54%	91.08%	73.42%	95.15%	90.91%	72.00%
7	e3_ensemble_sweep_k4_m1	–	0.058	m1	97.81%	99.19%	94.68%	100.00%	99.12%	98.41%	97.69%	90.98%	73.22%	96.08%	91.53%	72.00%
8	e3_ensemble_sweep_k5_m1	–	0.111	m1	98.24%	99.25%	96.32%	100.00%	99.12%	98.41%	97.97%	92.17%	78.95%	96.12%	91.53%	83.33%
9	e3_ensemble_sweep_k6_m1	–	0.087	m1	98.24%	99.25%	96.32%	100.00%	99.12%	98.41%	97.91%	91.76%	76.49%	96.12%	91.53%	86.96%
10	e3_ensemble_sweep_k7_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên mọi phương pháp qua bộ lọc · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1	–	0.087	m1	98.24%	99.25%	96.32%	100.00%	99.12%	98.41%	97.91%	91.76%	76.49%	96.12%	91.53%	86.96%
11	e3_ensemble_sweep_k8_m1	–	0.087	m1	98.24%	99.25%	96.32%	100.00%	99.12%	98.41%	97.91%	91.76%	76.49%	96.12%	91.53%	86.96%
12	e3_ensemble_top_geomean_m1	–	0.073	m1	97.88%	99.03%	95.40%	100.00%	99.12%	98.41%	97.44%	90.84%	72.63%	96.08%	92.74%	76.92%
13	e3_ensemble_top_mean_m1	–	0.073	m1	97.37%	98.87%	93.26%	100.00%	98.82%	98.41%	96.99%	89.78%	66.50%	97.03%	93.26%	76.92%
14	e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	97.19% ±0.09	98.69% ±0.23	93.90% ±1.63	99.87% ±0.22	99.02% ±0.17	98.15% ±0.92	96.38% ±0.64	90.40% ±1.17	67.00% ±6.50	94.92% ±1.18	89.32% ±2.04	70.56% ±4.19
15	e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1	intfloat/multilingual-e5-base	0.017	m1	97.22%	98.50%	93.58%	100.00%	98.82%	97.64%	95.57%	89.64%	63.16%	94.23%	86.21%	64.29%
16	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m1	jhu-clsp/mmBERT-base	0.155	m1	97.53%	98.76%	95.38%	100.00%	98.83%	98.41%	97.13%	89.87%	71.35%	92.61%	88.00%	74.07%
17	e2_qwen3_17b_base_genlabel_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.081	m1	96.82%	98.49%	92.02%	99.03%	98.82%	97.60%	96.71%	90.42%	65.31%	95.15%	89.39%	78.57%
18	e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-1.7B-Base	0.105	m1	96.89% ±0.19	98.50% ±0.08	93.00% ±0.56	99.55% ±0.40	98.97% ±0.15	97.61% ±0.79	96.83% ±0.47	90.60% ±0.37	66.78% ±0.59	95.28% ±1.35	91.31% ±0.43	85.83% ±3.76
19	e3_ensemble_top_greedy_m1	–	0.113	m1	97.80%	99.09%	94.84%	100.00%	98.97%	98.41%	97.31%	90.71%	69.63%	96.55%	92.66%	81.48%
20	e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m1	intfloat/multilingual-e5-large-instruct	0.017	m1	97.30% ±0.44	98.89% ±0.25	93.54% ±1.67	99.87% ±0.22	99.07% ±0.09	97.34% ±1.69	96.76% ±0.24	90.74% ±1.24	70.67% ±5.62	95.08% ±0.49	89.31% ±1.28	75.30% ±5.69
21	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1	jhu-clsp/mmBERT-base	0.108	m1	96.99% ±0.41	98.58% ±0.30	92.01% ±0.31	99.42% ±0.34	99.02% ±0.31	97.87% ±0.47	95.58% ±0.65	89.12% ±1.34	61.63% ±4.68	93.16% ±0.46	86.69% ±1.98	77.59% ±4.48
22	e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1	Qwen/Qwen3-4B-Base	0.056	m1	97.11% ±0.16	98.95% ±0.03	92.76% ±0.35	99.87% ±0.11	99.17% ±0.47	98.14% ±0.47	97.25% ±0.18	89.68% ±0.59	69.22% ±1.52	96.54% ±0.46	92.48% ±0.66	71.58% ±2.43
23	e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1	unsloth/gemma-3-1b-pt	0.045	m1	96.86% ±0.40	98.66% ±0.15	92.44% ±0.87	100.00% ±0.00	99.12% ±0.00	98.01% ±0.57	96.77% ±0.03	89.78% ±0.07	66.43% ±1.29	95.24% ±0.27	89.83% ±2.79	76.91% ±6.94
24	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepeval_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1, nhãn m1	vinai/phobert-base-v2	0.033	m1	97.11% ±0.05	98.62% ±0.14	92.17% ±0.31	99.87% ±0.11	98.48% ±0.23	98.41% ±0.00	96.41% ±0.18	88.91% ±0.67	64.17% ±1.59	95.30% ±1.06	88.67% ±1.88	69.53% ±7.59
25	e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1	unsloth/gemma-3-1b-pt	0.045	m1	97.13% ±0.35	98.67% ±0.21	92.30% ±1.29	99.74% ±0.11	99.17% ±0.17	98.42% ±0.79	96.36% ±0.53	88.50% ±0.37	64.34% ±5.34	95.63% ±0.74	89.57% ±2.96	74.66% ±7.59
26	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1	jhu-clsp/mmBERT-base	0.052	m1	97.63% ±0.62	98.67% ±0.20	93.05% ±0.89	99.74% ±0.29	98.83% ±0.29	98.16% ±0.91	96.43% ±0.19	89.52% ±1.55	66.25% ±2.67	93.08% ±1.00	88.32% ±2.63	71.41% ±4.77
27	e2_mmbert_base_cross_ce_synth1x_frozenemb_lr2e5_fbmatched_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	jhu-clsp/mmBERT-base	0.051	m1	97.53% ±0.31	98.81% ±0.11	93.87% ±0.86	99.68% ±0.29	98.97% ±0.39	98.42% ±0.01	96.84% ±0.29	90.45% ±1.08	67.83% ±2.59	95.16% ±0.75	89.01% ±0.58	77.49% ±3.74
28	e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.118	m1	97.24% ±0.24	98.54% ±0.00	93.63% ±0.60	99.94% ±0.11	99.27% ±0.29	98.14% ±0.47	96.91% ±0.28	89.11% ±1.23	67.06% ±4.33	94.84% ±1.14	88.75% ±1.21	76.05% ±8.16
29	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1	vinai/phobert-base-v2	0.008	m1	97.74%	98.81%	93.92%	99.81%	99.12%	94.21%	96.86%	89.52%	68.41%	94.12%	87.78%	28.57%
30	e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1	Qwen/Qwen3-4B-Base	0.162	m1	97.03%	98.70%	93.21%	99.62%	99.12%	98.41%	97.19%	88.02%	67.50%	96.59%	90.91%	80.00%
31	e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1	Qwen/Qwen3-8B-Base	0.170	m1	98.17%	99.08%	95.00%	99.42%	98.83%	98.41%	97.60%	91.04%	74.07%	96.08%	90.91%	78.57%
32	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_m1	vinai/phobert-base-v2	0.073	m1	97.40% ±0.24	98.67% ±0.03	93.63% ±0.90	99.68% ±0.22	98.83% ±0.15	98.42% ±1.38	97.00% ±0.09	88.32% ±0.20	68.33% ±0.80	95.59% ±0.49	89.26% ±2.95	73.34% ±2.28
33	e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1	unsloth/gemma-3-1b-pt	0.064	m1	97.57% ±0.19	98.78% ±0.06	92.98% ±0.50	99.74% ±0.29	99.12% ±0.15	98.68% ±0.46	96.91% ±0.03	89.38% ±0.53	68.90% ±0.45	95.53% ±0.82	89.42% ±3.29	71.15% ±3.33
34	e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-base	0.031	m1	96.77% ±0.23	98.01% ±0.32	92.29% ±1.28	99.55% ±0.48	99.32% ±0.09	97.87% ±0.47	96.44% ±0.42	89.68% ±1.05	67.68% ±3.73	93.89% ±0.54	88.12% ±1.68	66.67% ±5.02
35	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.061	m1	97.25% ±0.68	98.57% ±0.35	93.33% ±1.47	99.94% ±0.11	98.92% ±0.37	98.15% ±0.45	96.34% ±0.32	89.41% ±1.20	65.89% ±4.33	95.68% ±0.94	87.78% ±2.64	72.86% ±5.45
36	e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1	Qwen/Qwen3-4B-Base	0.131	m1	96.95% ±0.77	98.81% ±0.16	93.00% ±1.72	99.87% ±0.11	99.07% ±0.23	98.41% ±0.00	97.56% ±0.49	90.06% ±0.61	73.14% ±5.66	95.95% ±1.46	91.68% ±1.44	75.71% ±9.83
37	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.043	m1	96.71% ±0.44	98.22% ±0.42	92.45% ±1.27	99.87% ±0.11	98.68% ±0.30	96.77% ±0.83	96.54% ±0.19	88.85% ±0.32	66.53% ±1.38	95.59% ±0.43	89.89% ±2.25	73.58% ±6.68
38	e2_gemma3_1b_pt_cross_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.053	m1	97.15% ±0.38	98.63% ±0.22	92.21% ±0.44	100.00% ±0.00	98.97% ±0.00	97.35% ±0.50	96.75% ±0.36	89.57% ±0.94	67.92% ±2.37	95.16% ±0.24	88.67% ±1.26	73.16% ±4.63
39	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 3e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1	vinai/phobert-base-v2	0.037	m1	97.21% ±0.42	98.67% ±0.14	92.61% ±0.76	99.81% ±0.19	98.63% ±0.30	98.13% ±1.24	96.78% ±0.49	88.75% ±0.37	67.90% ±2.17	94.30% ±1.34	90.50% ±1.05	82.40% ±2.97
40	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m1	vinai/phobert-base-v2	0.085	m1	97.36% ±0.66	98.69% ±0.40	94.09% ±1.37	99.81% ±0.00	98.73% ±0.37	98.68% ±0.93	97.08% ±0.25	89.34% ±1.03	71.29% ±1.21	94.53% ±0.26	89.36% ±1.50	77.16% ±6.99
41	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_m1	vinai/phobert-base-v2	0.028	m1	97.14% ±0.40	98.56% ±0.38	92.38% ±0.52	99.87% ±0.22	98.73% ±0.45	98.68% ±0.46	96.76% ±0.38	88.36% ±0.52	65.73% ±1.19	95.87% ±0.72	89.27% ±0.98	68.48% ±13.47
42	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1	jhu-clsp/mmBERT-base	0.082	m1	97.36% ±0.21	98.61% ±0.54	93.03% ±1.24	99.74% ±0.11	99.02% ±0.08	97.88% ±0.46	96.69% ±0.24	88.93% ±1.65	67.37% ±4.94	94.37% ±0.22	86.73% ±2.27	68.29% ±1.41
43	e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.057	m1	96.95% ±0.27	98.67% ±0.23	92.72% ±0.93	99.55% ±0.22	99.02% ±0.09	98.41% ±0.00	96.47% ±0.17	89.30% ±0.95	63.89% ±1.43	94.82% ±0.40	88.03% ±0.85	72.92% ±5.08
44	e2_qwen3_17b_base_cross_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.101	m1	96.77% ±0.47	98.64% ±0.03	92.62% ±1.21	99.55% ±0.45	99.12% ±0.26	97.60% ±0.00	96.45% ±0.18	89.53% ±0.25	66.21% ±2.51	95.88% ±1.87	90.93% ±0.97	71.49% ±6.27
45	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m1	vinai/phobert-base-v2	0.052	m1	97.04% ±0.23	98.56% ±0.12	92.13% ±0.32	99.87% ±0.11	98.68% ±0.15	96.77% ±1.46	96.70% ±0.28	88.95% ±0.80	67.51% ±2.90	95.57% ±0.49	90.17% ±0.73	74.78% ±4.67
46	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1	vinai/phobert-base-v2	0.077	m1	96.74% ±0.14	98.34% ±0.15	92.73% ±1.15	99.87% ±0.11	99.03% ±0.30	97.87% ±0.47	97.02% ±0.42	89.08% ±1.68	69.66% ±5.81	94.98% ±0.70	91.20% ±0.63	78.73% ±2.20
47	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	vinai/phobert-base-v2	0.058	m1	96.71% ±0.22	98.26% ±0.20	92.37% ±0.72	99.81% ±0.19	98.48% ±0.17	97.32% ±1.27	96.65% ±0.16	88.60% ±0.13	68.40% ±1.72	95.22% ±0.76	89.12% ±2.35	69.89% ±6.54
48	e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1	unsloth/gemma-3-1b-pt	0.056	m1	97.30% ±0.13	98.63% ±0.22	92.54% ±0.69	99.87% ±0.22	99.07% ±0.09	98.42% ±0.01	96.85% ±0.28	88.79% ±0.74	66.21% ±3.02	94.10% ±0.47	88.94% ±2.49	74.72% ±5.03
49	e2_gemma3_270m_cross_ce_real_lora_m1gemma-3-270m · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn m1	unsloth/gemma-3-270m	0.098	m1	96.61% ±0.39	98.26% ±0.04	93.26% ±1.57	99.74% ±0.45	98.82% ±0.39	98.43% ±0.77	96.12% ±0.32	87.52% ±0.55	64.11% ±1.63	93.27% ±0.29	86.53% ±3.71	66.71% ±4.23
50	e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-large-instruct	0.029	m1	97.36% ±0.44	98.69% ±0.28	93.91% ±1.65	99.81% ±0.00	99.02% ±0.09	97.88% ±0.46	96.95% ±0.30	90.41% ±0.67	68.84% ±0.77	96.07% ±0.97	88.70% ±1.47	71.24% ±1.59
51	e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.034	m1	96.94% ±0.19	98.47% ±0.22	91.96% ±0.68	99.87% ±0.11	98.97% ±0.14	97.87% ±0.47	95.91% ±0.95	89.79% ±0.90	63.87% ±2.52	94.78% ±1.37	88.42% ±0.65	79.39% ±7.59
52	e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.051	m1	96.92% ±0.05	98.55% ±0.09	92.48% ±0.65	99.74% ±0.29	98.73% ±0.22	97.60% ±0.00	96.14% ±0.13	90.09% ±0.98	65.73% ±0.67	94.37% ±2.42	87.06% ±1.92	71.96% ±15.22
53	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1	jhu-clsp/mmBERT-base	0.040	m1	97.44% ±0.47	98.35% ±0.22	92.60% ±0.36	99.87% ±0.11	98.98% ±0.15	97.61% ±0.82	96.31% ±0.62	89.36% ±1.23	64.05% ±3.61	94.84% ±1.08	86.93% ±4.23	66.91% ±4.97
54	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1	jhu-clsp/mmBERT-base	0.057	m1	96.90% ±0.10	98.40% ±0.22	91.79% ±0.42	99.29% ±0.49	98.93% ±0.30	98.15% ±0.45	96.55% ±0.71	89.32% ±0.18	66.63% ±3.42	95.82% ±1.27	88.58% ±1.60	71.17% ±7.64
55	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.070	m1	97.52%	98.77%	94.25%	99.81%	98.97%	98.41%	96.80%	90.54%	70.30%	93.66%	88.00%	75.86%
56	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.047	m1	97.62% ±0.28	98.44% ±0.47	93.48% ±0.65	99.94% ±0.11	99.12% ±0.15	97.88% ±1.22	96.46% ±0.37	89.54% ±0.67	65.78% ±1.34	94.54% ±1.01	88.57% ±1.69	71.48% ±10.61
57	e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1	Qwen/Qwen3-1.7B-Base	0.095	m1	97.07% ±0.22	98.58% ±0.06	93.66% ±0.52	99.29% ±0.62	98.87% ±0.09	97.32% ±0.48	97.19% ±0.11	90.22% ±1.38	68.12% ±2.72	96.64% ±0.79	91.24% ±0.67	81.38% ±4.54
58	e2_xlmr_base_khonly_ce_real_frozenemb_m1XLM-R base · chỉ câu KH · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1	FacebookAI/xlm-roberta-base	0.012	m1	95.94%	97.74%	92.13%	100.00%	98.98%	96.77%	95.89%	88.39%	66.33%	92.45%	90.00%	64.00%
59	e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1	unsloth/gemma-3-1b-pt	0.042	m1	97.03% ±0.07	98.43% ±0.10	91.90% ±0.46	99.74% ±0.11	98.77% ±0.17	97.88% ±1.22	96.36% ±0.34	89.56% ±0.66	65.56% ±1.88	94.38% ±0.45	87.02% ±1.48	79.29% ±4.89
60	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m1	vinai/phobert-base-v2	0.024	m1	96.66% ±0.06	98.61% ±0.25	91.61% ±0.16	99.87% ±0.11	98.68% ±0.00	97.32% ±1.26	96.58% ±0.14	88.44% ±0.71	65.19% ±0.43	95.95% ±0.52	89.60% ±3.82	76.31% ±3.64
61	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m1	vinai/phobert-base-v2	0.031	m1	97.11% ±0.12	98.47% ±0.32	92.88% ±0.63	99.81% ±0.19	98.68% ±0.00	97.01% ±2.42	96.82% ±0.38	88.54% ±1.57	66.63% ±3.82	95.25% ±0.73	88.65% ±3.23	75.65% ±5.05
62	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m1	vinai/phobert-base-v2	0.031	m1	96.78% ±0.22	98.38% ±0.24	91.90% ±0.69	99.81% ±0.19	98.53% ±0.15	97.60% ±0.82	96.77% ±0.43	89.19% ±0.47	68.39% ±3.53	94.31% ±0.24	88.45% ±2.26	77.00% ±3.80
63	e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1	unsloth/gemma-3-1b-pt	0.096	m1	97.07% ±0.33	98.62% ±0.11	92.94% ±0.84	99.81% ±0.33	99.02% ±0.23	98.43% ±1.35	96.73% ±0.39	88.81% ±0.92	67.39% ±4.54	95.36% ±1.55	89.94% ±2.05	49.97% ±21.22
64	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1	jhu-clsp/mmBERT-base	0.061	m1	96.84% ±0.72	98.36% ±0.27	91.78% ±0.71	99.68% ±0.11	99.12% ±0.25	97.61% ±0.82	95.61% ±0.31	89.47% ±0.87	62.47% ±1.97	94.55% ±1.25	88.45% ±1.79	70.04% ±11.51
65	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.017	m1	97.46% ±0.37	98.62% ±0.16	93.95% ±0.35	99.94% ±0.11	99.02% ±0.09	95.94% ±1.49	96.81% ±0.07	89.78% ±0.50	68.55% ±0.92	95.87% ±0.26	88.85% ±1.47	70.12% ±3.51
66	e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1	vinai/phobert-base-v2	0.026	m1	97.47% ±0.24	98.44% ±0.16	94.80% ±0.96	100.00% ±0.00	98.92% ±0.08	96.82% ±0.05	97.02% ±0.20	89.76% ±0.51	70.75% ±3.75	95.43% ±1.00	88.65% ±1.18	72.62% ±5.19
67	e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.065	m1	96.64% ±0.34	98.56% ±0.13	92.00% ±0.57	99.87% ±0.11	98.78% ±0.31	97.61% ±0.79	96.72% ±0.13	88.75% ±0.72	67.75% ±1.81	94.29% ±0.99	89.52% ±0.82	73.00% ±5.87
68	e2_xlmr_base_cross_logitadjust_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.018	m1	96.61% ±0.24	98.11% ±0.05	91.94% ±0.31	99.62% ±0.33	99.02% ±0.09	97.61% ±0.82	96.14% ±0.27	87.44% ±0.74	63.23% ±2.11	94.09% ±1.39	88.49% ±1.47	69.23% ±0.00
69	e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1	intfloat/multilingual-e5-large-instruct	0.026	m1	97.35% ±0.55	98.76% ±0.09	93.89% ±1.61	99.94% ±0.11	99.02% ±0.09	97.88% ±0.92	96.82% ±0.45	91.40% ±0.48	71.28% ±4.58	95.07% ±0.42	90.32% ±0.63	73.19% ±1.53
70	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1	jhu-clsp/mmBERT-base	0.072	m1	97.65% ±0.33	98.37% ±0.37	94.06% ±0.19	99.74% ±0.29	99.02% ±0.09	98.41% ±0.81	96.72% ±0.42	89.81% ±1.04	67.76% ±2.08	94.68% ±0.27	86.84% ±4.27	69.39% ±9.12
71	e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1	jhu-clsp/mmBERT-base	0.082	m1	96.90% ±0.28	98.53% ±0.11	92.45% ±0.23	99.55% ±0.40	98.68% ±0.14	98.15% ±0.45	96.57% ±0.03	90.00% ±0.60	67.77% ±1.03	94.76% ±0.87	86.75% ±3.93	68.25% ±7.90
72	e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1	Qwen/Qwen3-1.7B-Base	0.067	m1	96.45%	98.38%	92.11%	99.81%	99.12%	97.60%	95.88%	88.64%	61.18%	95.57%	89.77%	74.07%
73	e2_visobert_cross_ce_real_frozenemb_lr4e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1	uitnlp/visobert	0.037	m1	97.07%	98.12%	92.98%	99.62%	98.98%	96.83%	95.01%	87.27%	56.10%	92.54%	88.14%	72.00%
74	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1	FacebookAI/xlm-roberta-base	0.027	m1	96.52%	98.17%	92.45%	99.81%	99.12%	96.77%	95.50%	89.56%	62.00%	94.69%	88.76%	72.00%
75	e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m1mmBERT base · dual encoder (câu bot và câu KH mã hoá riêng, pooling mean), full fine-tune, đóng băng embedding · fusion [u; v; \|u−v\|; u·v], head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1	jhu-clsp/mmBERT-base	0.057	m1	96.82% ±0.28	98.17% ±0.34	92.52% ±0.49	99.94% ±0.11	98.92% ±0.17	97.89% ±0.50	96.41% ±0.29	89.75% ±1.53	66.08% ±3.76	96.46% ±0.47	90.22% ±1.52	73.73% ±3.28
76	e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.206	m1	97.25% ±0.52	98.39% ±0.10	94.61% ±1.42	99.87% ±0.11	98.63% ±0.23	98.68% ±0.46	96.96% ±0.42	89.98% ±1.80	74.91% ±7.57	94.96% ±0.71	90.04% ±1.05	75.33% ±10.98
77	e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.034	m1	96.72% ±0.18	98.20% ±0.18	92.46% ±0.59	100.00% ±0.00	98.72% ±0.34	97.06% ±1.26	96.94% ±0.12	89.39% ±0.48	70.60% ±2.41	95.32% ±0.55	89.58% ±0.51	73.81% ±6.30
78	e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.052	m1	96.89% ±0.27	98.38% ±0.33	93.04% ±1.32	100.00% ±0.00	98.58% ±0.09	98.15% ±0.45	97.01% ±0.43	89.16% ±0.79	69.66% ±3.80	95.91% ±1.11	87.62% ±3.03	69.00% ±6.46
79	e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_m1	vinai/phobert-base-v2	0.033	m1	96.70% ±0.63	98.33% ±0.28	92.28% ±1.15	99.94% ±0.11	98.58% ±0.08	97.32% ±0.48	96.88% ±0.27	89.00% ±0.98	69.57% ±2.84	95.11% ±1.23	87.35% ±2.90	70.69% ±8.52
80	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1	jhu-clsp/mmBERT-base	0.144	m1	97.12% ±0.71	98.44% ±0.43	93.10% ±1.04	99.36% ±0.40	98.83% ±0.29	97.87% ±0.95	96.32% ±0.21	89.85% ±0.56	66.13% ±2.75	94.08% ±0.49	88.59% ±0.98	69.11% ±2.38
81	e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1	jhu-clsp/mmBERT-base	0.082	m1	97.45% ±0.50	98.47% ±0.32	93.34% ±2.22	99.94% ±0.11	99.12% ±0.15	98.15% ±0.45	96.41% ±0.71	89.62% ±0.63	66.28% ±4.39	95.00% ±0.95	87.86% ±1.59	76.08% ±11.13
82	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1	jhu-clsp/mmBERT-base	0.108	m1	97.15% ±0.07	98.58% ±0.11	92.69% ±1.15	99.48% ±0.56	98.98% ±0.25	97.60% ±0.00	96.93% ±0.13	90.16% ±0.89	71.20% ±3.14	95.35% ±1.19	87.22% ±1.46	72.53% ±3.96
83	e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1	Qwen/Qwen3-0.6B-Base	0.075	m1	96.83% ±0.37	98.40% ±0.12	92.62% ±1.56	100.00% ±0.00	98.43% ±0.09	97.93% ±0.44	96.43% ±0.52	89.02% ±1.09	65.75% ±5.65	94.66% ±0.92	92.53% ±1.97	72.95% ±1.36
84	e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1	Qwen/Qwen3-0.6B-Base	0.077	m1	96.45% ±0.18	97.90% ±0.25	92.02% ±0.64	98.77% ±0.11	98.87% ±0.45	97.60% ±0.00	95.14% ±0.21	88.12% ±1.81	60.66% ±2.95	95.65% ±1.05	91.13% ±1.54	75.64% ±8.41
85	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_m1	vinai/phobert-base-v2	0.093	m1	96.59% ±0.40	98.20% ±0.39	92.19% ±0.85	99.49% ±0.29	98.63% ±0.22	98.42% ±0.01	96.73% ±0.57	90.23% ±1.06	69.83% ±0.95	95.04% ±1.45	88.81% ±1.37	77.28% ±1.98
86	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m1	vinai/phobert-base-v2	0.104	m1	97.11% ±0.25	98.43% ±0.14	92.85% ±1.30	99.68% ±0.11	98.83% ±0.30	97.87% ±0.47	96.87% ±0.31	89.38% ±0.92	71.41% ±5.94	95.15% ±1.26	87.43% ±2.20	78.02% ±3.42
87	e2_xlmr_base_cross_ce_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.044	m1	96.51% ±0.41	98.25% ±0.31	92.20% ±1.02	99.87% ±0.11	99.22% ±0.22	95.95% ±0.85	96.19% ±0.32	88.07% ±2.17	64.64% ±5.21	93.22% ±0.52	89.02% ±0.37	69.09% ±5.23
88	e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1	FacebookAI/xlm-roberta-base	0.019	m1	96.65% ±0.24	98.25% ±0.06	92.55% ±0.85	99.81% ±0.00	99.12% ±0.15	97.90% ±0.46	96.18% ±0.25	88.12% ±0.75	66.66% ±3.81	94.06% ±0.25	88.35% ±0.68	74.00% ±3.97
89	e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m1	vinai/phobert-base-v2	0.049	m1	96.47% ±0.52	98.26% ±0.33	91.55% ±1.05	99.49% ±0.55	98.68% ±0.30	97.32% ±1.26	96.61% ±0.17	88.63% ±0.93	66.96% ±0.69	95.41% ±0.51	89.01% ±1.44	68.25% ±2.75
90	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.159	m1	96.37% ±0.22	97.95% ±0.11	92.00% ±0.78	99.43% ±0.68	98.63% ±0.23	95.72% ±0.91	96.66% ±0.21	88.74% ±1.26	69.99% ±6.38	95.30% ±1.01	88.20% ±2.45	60.65% ±22.95
91	e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-4B	0.061	m1	94.82% ±1.69	97.27% ±1.59	90.99% ±0.31	98.34% ±0.70	98.46% ±0.31	98.43% ±0.02	95.50% ±0.10	86.68% ±1.10	63.49% ±1.67	93.26% ±5.15	84.63% ±0.41	71.46% ±14.64
92	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1	vinai/phobert-base-v2	0.025	m1	97.46%	97.84%	91.67%	99.81%	98.53%	94.21%	96.88%	88.97%	68.29%	94.63%	85.39%	25.00%
93	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.152	m1	95.59% ±0.34	97.66% ±0.23	92.30% ±0.68	99.49% ±0.40	98.78% ±0.09	97.90% ±0.46	95.82% ±0.32	87.61% ±1.28	65.34% ±3.51	92.69% ±0.71	84.80% ±1.62	69.76% ±4.64
94	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m1	vinai/phobert-base-v2	0.035	m1	96.64% ±0.17	98.54% ±0.14	91.53% ±1.47	99.37% ±0.61	98.54% ±0.15	97.07% ±0.92	96.65% ±0.48	89.24% ±1.17	68.27% ±4.26	94.96% ±1.02	89.30% ±1.04	73.81% ±2.06
95	e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1	intfloat/multilingual-e5-base	0.052	m1	95.65%	97.69%	91.12%	98.86%	99.12%	96.00%	95.80%	89.32%	63.50%	91.87%	88.37%	33.33%
96	e2_visobert_cross_ce_real_frozenemb_lr2e5_m1	uitnlp/visobert	0.027	m1	96.59%	97.37%	91.78%	99.81%	98.83%	96.00%	95.19%	85.79%	57.86%	94.00%	88.89%	58.33%
97	e2_visobert_cross_ce_real_frozenemb_lr1e5_m1	uitnlp/visobert	0.186	m1	95.07%	96.52%	90.80%	99.24%	98.54%	96.77%	94.98%	86.27%	59.08%	93.53%	82.56%	61.54%
98	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1	FacebookAI/xlm-roberta-base	0.161	m1	95.76%	97.57%	90.80%	98.86%	99.27%	95.08%	95.58%	88.42%	66.49%	92.38%	86.21%	51.85%
99	e1_tfidf_hybrid_qwen3emb_lr_m1TF-IDF câu KH (char 1-5, word 1-2-gram, giữ từ 1 ký tự, thêm kênh char bỏ dấu) + câu bot (char 2-4) + tương tác loại câu bot × từ câu KH · vector đóng băng Qwen3-Embedding-0.6B (cặp bot + KH, pooling last_token), trọng số dense 1 · LogReg · train thật, nhãn m1	Qwen/Qwen3-Embedding-0.6B	0.149	m1	95.17%	96.46%	89.87%	99.23%	98.53%	99.21%	95.66%	84.67%	60.76%	91.28%	85.88%	69.57%
100	e1_tfidf_svmcal_ce_synth1x_fbmatched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1	TF-IDF	0.207	m1	94.68%	96.37%	89.71%	99.04%	98.53%	97.60%	95.92%	87.30%	67.84%	93.07%	86.39%	72.73%
101	e1_tfidf_svmcal_ce_synth1x_nofb_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent omi_intent_repeat, callback, relationship, negate, affirm), nhãn m1	TF-IDF	0.164	m1	94.59%	96.31%	89.19%	99.24%	98.53%	97.60%	95.98%	87.25%	65.90%	93.07%	86.39%	69.57%
102	e1_tfidf_svmcal_ce_synth1x_noqparticle_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm q_particle), nhãn m1	TF-IDF	0.280	m1	94.46%	96.16%	88.97%	99.23%	98.53%	97.60%	95.32%	86.40%	60.91%	92.61%	86.23%	80.00%
103	e1_tfidf_svmcal_ce_real_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn m1	TF-IDF	0.178	m1	94.60%	96.43%	89.11%	99.43%	98.38%	97.60%	96.04%	87.13%	66.67%	93.07%	86.39%	69.57%
104	e1_tfidf_svmcal_ce_synth1x_nofiller_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm filler_only), nhãn m1	TF-IDF	0.267	m1	94.62%	96.25%	88.42%	99.42%	98.53%	97.60%	95.48%	86.19%	61.85%	92.61%	85.54%	78.26%
105	e1_tfidf_svmcal_ce_synth1x_noofftopic_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm prefix_offtopic), nhãn m1	TF-IDF	0.298	m1	94.39%	96.21%	88.59%	99.62%	98.53%	97.60%	95.33%	86.58%	62.02%	92.61%	86.39%	76.92%
106	e2_dual_qwen3emb06b_concat_ce_real_frozen_m1	Qwen/Qwen3-Embedding-0.6B	0.227	m1	94.56% ±0.08	96.66% ±0.17	89.27% ±0.45	98.22% ±0.49	96.96% ±0.06	97.88% ±0.46	94.40% ±0.49	83.82% ±0.45	60.36% ±2.35	93.48% ±1.41	76.59% ±3.88	77.23% ±4.17
107	e1_tfidf_svmcal_ce_synth1x_base_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1, nhãn m1	TF-IDF	0.282	m1	94.40%	96.47%	88.37%	99.23%	98.53%	97.60%	95.29%	86.16%	60.56%	92.61%	85.71%	76.92%
108	e1_tfidf_svmcal_ce_synth1x_matched_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (khớp tỉ lệ intent của train thật), nhãn m1	TF-IDF	0.190	m1	94.45%	96.32%	88.72%	99.04%	98.38%	97.60%	96.01%	87.05%	66.67%	93.07%	85.71%	69.57%
109	e1_tfidf_svmcal_ce_synth1x_nomoney_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (bỏ nhóm money), nhãn m1	TF-IDF	0.310	m1	94.56%	96.26%	89.06%	99.23%	98.53%	97.60%	95.39%	86.61%	62.86%	92.61%	87.06%	80.00%
110	e1_tfidf_svmcal_ce_synth1x_fbonly_m1TF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật + synth ×1 (chỉ intent nlu_fallback), nhãn m1	TF-IDF	0.356	m1	94.83%	96.38%	89.55%	99.62%	98.38%	97.60%	94.80%	86.61%	58.88%	92.61%	86.39%	76.92%
–	e2_bamibert_cross_ce_real_frozenemb_lr1e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_bamibert_cross_ce_real_frozenemb_lr2e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_bamibert_cross_ce_real_frozenemb_lr4e5_m1	–	–	m1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e3_fbverifier_devtop1_familygeomean_fbf1_m1	–	hard	m1	98.18%	98.93%	95.41%	100.00%	99.12%	97.67%	98.07%	92.94%	83.74%	95.65%	91.11%	75.86%
–	e3_fbverifier_devtop1_familygeomean_macrof1_m1	–	hard	m1	98.18%	98.93%	95.41%	100.00%	99.12%	97.67%	98.07%	92.94%	83.74%	95.65%	91.11%	75.86%
1	e3_ensemble_mean_all	–	0.065	v1	97.86%	98.51%	92.83%	98.88%	99.14%	96.00%	97.05%	89.84%	69.00%	92.66%	89.01%	71.43%
2	e3_ensemble_greedy_all	–	0.070	v1	97.56%	98.67%	92.31%	98.50%	99.14%	96.88%	96.96%	89.92%	68.50%	92.24%	88.40%	71.43%
3	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5	–	0.023	v1	97.60%	98.40%	93.62%	99.07%	98.99%	96.06%	97.00%	90.51%	71.16%	90.99%	86.03%	60.61%
4	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1	vinai/phobert-base-v2	0.011	v1	97.35%	98.77%	92.92%	98.69%	98.13%	96.83%	96.86%	88.45%	68.39%	93.52%	89.62%	61.54%
5	e2_xlmr_base_cross_wcesqrt_real_frozenemb	FacebookAI/xlm-roberta-base	0.021	v1	96.91% ±0.37	98.03% ±0.22	91.27% ±0.13	99.07% ±0.19	98.95% ±0.22	96.55% ±0.48	96.74% ±0.04	89.46% ±0.69	68.81% ±1.73	92.71% ±1.14	88.13% ±1.36	69.11% ±4.90
6	e3_prior_em_mmbert_bottype	–	0.082	v1	97.51%	98.51%	91.76%	98.50%	99.00%	96.83%	96.71%	75.82%	72.18%	90.57%	43.06%	66.67%
7	e3_prior_em_mmbert_global	–	0.082	v1	97.51%	98.51%	91.76%	98.50%	99.00%	96.83%	96.76%	88.87%	69.77%	91.67%	85.71%	80.00%
8	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5	jhu-clsp/mmBERT-base	0.061	v1	97.52% ±0.16	98.44% ±0.33	92.13% ±0.46	98.50% ±0.19	98.86% ±0.65	96.27% ±0.96	97.05% ±0.12	89.61% ±1.14	68.41% ±2.50	91.67% ±1.13	88.95% ±2.07	79.06% ±3.47
9	e2_xlmr_base_cross_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	97.15% ±0.54	98.26% ±0.54	93.21% ±1.52	99.07% ±0.32	98.57% ±0.14	95.00% ±1.14	96.88% ±0.09	89.87% ±0.73	71.41% ±0.44	92.38% ±0.49	87.47% ±0.61	71.67% ±5.13
10	e2_xlmr_base_cross_logitadjust_real_frozenemb	FacebookAI/xlm-roberta-base	0.035	v1	97.21% ±0.23	98.30% ±0.36	91.68% ±0.68	99.13% ±0.11	99.14% ±0.25	96.08% ±0.82	96.86% ±0.24	89.47% ±0.38	68.13% ±1.52	93.29% ±1.14	89.58% ±1.46	65.30% ±1.19
11	e2_qwen3_06b_base_cross_ce_real_lora	Qwen/Qwen3-0.6B-Base	0.098	v1	97.16%	97.85%	93.07%	98.32%	98.56%	97.64%	96.59%	88.93%	71.28%	92.59%	88.04%	71.43%
12	e3_ensemble_stack_all	–	0.088	v1	96.85%	98.72%	90.75%	98.31%	99.14%	94.40%	96.50%	90.42%	69.59%	92.45%	88.27%	74.07%
13	e2_qwen3_17b_base_cross_ce_real_lora	Qwen/Qwen3-1.7B-Base	0.083	v1	97.02% ±0.20	98.13% ±0.07	91.06% ±0.75	98.04% ±0.95	98.71% ±0.43	95.76% ±1.24	96.79% ±0.04	89.24% ±0.37	67.78% ±1.49	92.86% ±0.23	89.57% ±1.08	77.14% ±4.95
14	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5	vinai/phobert-base-v2	0.022	v1	96.92% ±0.17	98.21% ±0.21	93.39% ±1.17	98.88% ±0.18	98.27% ±0.14	94.60% ±1.82	97.01% ±0.10	89.86% ±0.32	70.55% ±0.40	92.79% ±0.86	88.37% ±1.90	69.29% ±7.44
15	e2_me5_base_cross_ce_real_frozenemb_lr1e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	intfloat/multilingual-e5-base	0.088	v1	96.62%	97.95%	92.53%	98.88%	98.57%	95.24%	96.47%	89.18%	70.98%	90.00%	88.40%	62.07%
16	e2_me5_base_cross_ce_real_frozenemb_lr4e5	–	0.028	v1	97.08%	98.21%	91.04%	98.69%	99.00%	95.24%	96.94%	89.50%	71.43%	92.66%	87.43%	76.92%
17	e2_xlmr_base_khonly_ce_real_frozenemb	FacebookAI/xlm-roberta-base	0.076	v1	96.31% ±0.48	97.53% ±0.31	92.74% ±0.96	98.57% ±0.29	98.57% ±0.25	96.30% ±0.46	96.64% ±0.20	89.54% ±1.07	70.84% ±1.93	91.79% ±0.49	89.20% ±1.47	73.27% ±4.80
18	e2_me5_base_cross_ce_real_frozenemb_lr2e5	intfloat/multilingual-e5-base	0.056	v1	96.90% ±0.22	98.15% ±0.15	92.39% ±0.25	98.82% ±0.10	98.95% ±0.08	96.55% ±2.03	96.85% ±0.12	90.17% ±0.33	70.92% ±1.25	91.43% ±0.42	88.49% ±1.38	70.15% ±6.37
19	e2_gemma3_270m_cross_ce_real_lora	unsloth/gemma-3-270m	0.116	v1	96.62% ±0.13	97.95% ±0.14	91.83% ±1.49	98.31% ±0.34	98.42% ±0.15	97.38% ±0.44	96.60% ±0.27	89.36% ±0.34	68.93% ±1.54	91.22% ±0.38	86.29% ±1.04	68.97% ±3.86
20	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen	Qwen/Qwen3-Embedding-0.6B	0.114	v1	96.17% ±0.21	97.45% ±0.03	91.07% ±0.55	98.62% ±0.22	98.42% ±0.15	97.10% ±0.47	96.37% ±0.05	88.63% ±0.28	67.53% ±0.61	90.41% ±0.74	89.42% ±1.19	73.49% ±4.11
21	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5	–	0.075	v1	97.53%	98.36%	91.44%	98.69%	99.28%	96.00%	97.12%	90.15%	68.89%	92.52%	89.01%	71.43%
22	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5	–	0.026	v1	96.65%	97.58%	91.27%	97.97%	98.41%	89.83%	96.68%	88.04%	69.04%	91.32%	87.91%	28.57%
23	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	FacebookAI/xlm-roberta-base	0.123	v1	96.23%	97.69%	90.94%	98.88%	98.27%	87.60%	96.52%	90.46%	70.11%	91.96%	86.36%	21.05%
24	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5	jhu-clsp/mmBERT-base	0.120	v1	97.12% ±0.21	98.33% ±0.04	90.60% ±0.02	98.12% ±0.26	98.50% ±0.51	96.50% ±0.53	96.75% ±0.05	89.96% ±0.69	70.64% ±0.07	93.03% ±0.61	88.06% ±0.21	65.73% ±4.57
25	e2_visobert_cross_ce_real_frozenemb_lr2e5	–	0.093	v1	95.71%	96.44%	90.22%	97.74%	99.14%	95.24%	96.48%	89.43%	68.63%	91.16%	85.87%	66.67%
26	e2_visobert_cross_ce_real_frozenemb_lr4e5	–	0.066	v1	96.33%	97.29%	91.18%	97.92%	98.58%	95.24%	96.29%	89.46%	67.18%	92.09%	86.96%	75.86%
27	e1_tfidf_svmcal_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · SVM calibrated · train thật, nhãn v1	TF-IDF	0.123	v1	95.17%	95.49%	89.49%	98.13%	98.25%	96.88%	96.57%	87.88%	67.49%	91.24%	87.36%	76.92%
28	e1_tfidf_logreg_balanced_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg cân bằng lớp · train thật, nhãn v1	TF-IDF	0.256	v1	94.94%	95.53%	90.11%	98.13%	98.56%	97.67%	96.03%	84.93%	63.31%	88.69%	88.30%	68.75%
29	e1_tfidf_logreg_ce_real_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật, nhãn v1	TF-IDF	0.125	v1	95.11%	95.69%	89.78%	98.32%	98.41%	96.06%	96.23%	86.51%	64.59%	90.23%	84.75%	74.07%
30	e2_visobert_cross_ce_real_frozenemb_lr1e5	–	0.232	v1	94.32%	95.91%	90.33%	97.23%	97.72%	92.91%	96.40%	88.60%	67.04%	89.09%	84.62%	62.50%
31	e1_tfidf_logreg_ce_synth1x_baseTF-IDF câu KH (char 1-5, word 1-2-gram) + câu bot (char 2-4) · LogReg · train thật + synth ×1, nhãn v1	TF-IDF	0.234	v1	94.74%	95.48%	88.97%	98.32%	98.11%	95.24%	96.25%	86.96%	66.50%	90.74%	85.23%	71.43%
–	e2_mdeberta_cross_ce_real_frozenemb_lr1e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_mdeberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	0.00%	52.16%	0.00%	0.00%	0.00%	0.00%	0.00%	36.70%	0.00%	0.00%	0.00%	0.00%
–	e2_videberta_cross_ce_real_frozenemb_lr1e5	–	argmax	v1	0.00%	52.16%	0.00%	0.00%	0.00%	0.00%	0.00%	36.70%	0.00%	0.00%	0.00%	0.00%
–	e2_videberta_cross_ce_real_frozenemb_lr2e5	–	argmax	v1	0.00%	52.16%	0.00%	0.00%	0.00%	0.00%	0.00%	36.70%	0.00%	0.00%	0.00%	0.00%
–	e2_videberta_cross_ce_real_frozenemb_lr2e5_fp32	–	–	v1	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
–	e2_videberta_cross_ce_real_frozenemb_lr4e5	–	argmax	v1	0.00%	52.16%	0.00%	0.00%	0.00%	0.00%	0.00%	36.70%	0.00%	0.00%	0.00%	0.00%
1	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_lv21	jhu-clsp/mmBERT-base	0.076	v2_1	97.61% ±0.23	98.31% ±0.38	91.68% ±1.08	98.04% ±0.50	99.04% ±0.16	97.11% ±0.46	96.82% ±0.18	90.15% ±0.56	69.73% ±2.33	90.45% ±0.80	89.04% ±2.07	74.52% ±4.59
2	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_lv21	vinai/phobert-base-v2	0.033	v2_1	96.63%	97.95%	94.57%	97.30%	98.27%	98.44%	96.88%	90.16%	72.46%	89.52%	84.15%	66.67%

ASR vs chép tay (chỉ để chẩn đoán)

Chép tay: cùng dòng và nhãn với split ASR tương ứng, câu KH thay bằng bản chép tay của QC (dòng không có bản chép tay bị bỏ). Production gửi chữ ASR nên số ở đây chỉ đo mức mất do lỗi ASR; không vào bảng xếp hạng, không chọn ngưỡng hay ensemble. Mỗi nhóm cột: ASR (trung bình các seed có cả hai split), chép tay (± độ lệch chuẩn qua seed), chênh = chép tay − ASR tính theo từng seed rồi lấy trung bình, đơn vị điểm phần trăm. Quy tắc giải mã mặc định; không tô ô tốt nhất.

STT	Model_Name	Data	clean seed · dòng ASR/chép tay	clean Macro_F1 ASR	clean Macro_F1 chép tay	clean Macro_F1 chênh	clean Macro_F1 −rep ASR	clean Macro_F1 −rep chép tay	clean Macro_F1 −rep chênh	clean FB recall ASR	clean FB recall chép tay	clean FB recall chênh	clean FB precision ASR	clean FB precision chép tay	clean FB precision chênh	reply-unseen seed · dòng ASR/chép tay	reply-unseen Macro_F1 ASR	reply-unseen Macro_F1 chép tay	reply-unseen Macro_F1 chênh	reply-unseen Macro_F1 −rep ASR	reply-unseen Macro_F1 −rep chép tay	reply-unseen Macro_F1 −rep chênh	reply-unseen FB recall ASR	reply-unseen FB recall chép tay	reply-unseen FB recall chênh	reply-unseen FB precision ASR	reply-unseen FB precision chép tay	reply-unseen FB precision chênh
14	e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m1	m1	3 · 2358.0/2358.0	84.53%	86.70% ±1.32	+2.2	87.32%	89.13% ±1.64	+1.8	94.62%	92.56% ±0.89	-2.1	51.43%	52.56% ±6.96	+1.1	3 · 1942.0/1942.0	78.50%	81.22% ±1.89	+2.7	86.97%	88.89% ±1.67	+1.9	95.08%	92.90% ±0.95	-2.2	50.78%	52.02% ±6.76	+1.2
15	e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1	m1	1 · 2358.0/2358.0	81.93%	85.99%	+4.1	85.46%	88.19%	+2.7	93.85%	93.08%	-0.8	46.92%	51.27%	+4.3	1 · 1942.0/1942.0	76.93%	81.42%	+4.5	84.82%	87.70%	+2.9	93.44%	92.62%	-0.8	45.60%	50.00%	+4.4
16	e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m1	m1	1 · 2358.0/2358.0	85.29%	87.51%	+2.2	87.53%	89.63%	+2.1	86.15%	84.62%	-1.5	60.22%	64.71%	+4.5	1 · 1942.0/1942.0	81.46%	83.94%	+2.5	87.09%	89.30%	+2.2	86.07%	84.43%	-1.6	58.99%	63.58%	+4.6
17	e2_qwen3_17b_base_genlabel_ce_real_lora_m1	m1	1 · 2358.0/2358.0	86.04%	88.36%	+2.3	87.53%	90.31%	+2.8	91.54%	91.54%	+0.0	51.97%	57.49%	+5.5	1 · 1942.0/1942.0	82.94%	85.36%	+2.4	87.03%	89.94%	+2.9	90.98%	90.98%	+0.0	50.45%	56.06%	+5.6
18	e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m1	m1	3 · 2358.0/2358.0	87.81%	89.47% ±0.31	+1.7	87.99%	89.91% ±0.46	+1.9	93.33%	91.28% ±1.18	-2.1	51.49%	55.46% ±0.84	+4.0	3 · 1942.0/1942.0	85.44%	87.32% ±0.37	+1.9	87.51%	89.53% ±0.50	+2.0	92.90%	90.71% ±1.25	-2.2	50.68%	54.54% ±1.29	+3.9
20	e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m1	m1	3 · 2358.0/2358.0	86.18%	88.64% ±0.35	+2.5	88.36%	90.54% ±1.35	+2.2	94.36%	92.82% ±0.44	-1.5	56.58%	62.15% ±8.85	+5.6	3 · 1942.0/1942.0	82.53%	85.35% ±0.13	+2.8	87.99%	90.31% ±1.37	+2.3	94.54%	92.90% ±0.47	-1.6	55.64%	61.38% ±8.74	+5.7
21	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1	m1	3 · 2358.0/2358.0	83.77%	86.26% ±2.47	+2.5	85.01%	87.58% ±1.87	+2.6	92.56%	91.54% ±5.55	-1.0	46.27%	51.11% ±8.58	+4.8	3 · 1942.0/1942.0	80.16%	82.98% ±3.75	+2.8	84.82%	87.61% ±1.84	+2.8	92.08%	90.98% ±5.91	-1.1	46.89%	52.22% ±8.31	+5.3
22	e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1	m1	3 · 2358.0/2358.0	86.01%	88.81% ±0.04	+2.8	88.90%	91.27% ±0.41	+2.4	95.38%	94.36% ±0.44	-1.0	54.20%	60.15% ±1.49	+6.0	3 · 1942.0/1942.0	81.84%	85.44% ±0.37	+3.6	88.46%	91.00% ±0.51	+2.5	95.08%	93.99% ±0.47	-1.1	53.35%	59.44% ±1.66	+6.1
23	e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1	m1	2 · 2358.0/2358.0	85.57%	87.89% ±0.40	+2.3	87.30%	89.39% ±1.03	+2.1	95.00%	93.08% ±0.00	-1.9	50.21%	55.53% ±1.80	+5.3	2 · 1942.0/1942.0	81.29%	84.00% ±1.05	+2.7	86.93%	89.14% ±1.09	+2.2	95.49%	93.44% ±0.00	-2.0	49.49%	54.84% ±1.86	+5.3
24	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepeval_m1	m1	3 · 2358.0/2358.0	83.55%	85.75% ±2.00	+2.2	86.35%	88.10% ±0.18	+1.7	91.79%	91.54% ±0.77	-0.3	48.47%	49.64% ±1.93	+1.2	3 · 1942.0/1942.0	77.76%	80.33% ±4.15	+2.6	86.10%	87.95% ±0.33	+1.9	91.80%	91.53% ±0.47	-0.3	48.25%	49.47% ±2.08	+1.2
25	e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1	m1	3 · 2358.0/2358.0	84.63%	86.85% ±2.90	+2.2	86.62%	88.11% ±1.66	+1.5	94.10%	91.79% ±0.44	-2.3	48.73%	51.24% ±5.42	+2.5	3 · 1942.0/1942.0	80.72%	84.06% ±4.81	+3.3	86.23%	87.92% ±1.88	+1.7	94.54%	92.35% ±0.47	-2.2	47.68%	50.67% ±5.94	+3.0
26	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1	m1	3 · 2358.0/2358.0	83.89%	85.91% ±0.61	+2.0	86.38%	88.49% ±0.54	+2.1	91.03%	88.72% ±1.60	-2.3	51.25%	54.75% ±3.40	+3.5	3 · 1942.0/1942.0	80.76%	82.80% ±0.47	+2.0	86.13%	88.34% ±0.52	+2.2	90.98%	88.52% ±1.42	-2.5	51.33%	54.93% ±2.61	+3.6
27	e2_mmbert_base_cross_ce_synth1x_frozenemb_lr2e5_fbmatched_m1	m1	3 · 2358.0/2358.0	85.91%	88.31% ±1.50	+2.4	87.60%	89.84% ±0.70	+2.2	91.79%	90.51% ±3.11	-1.3	53.27%	57.25% ±5.13	+4.0	3 · 1942.0/1942.0	82.53%	85.29% ±2.09	+2.8	87.23%	89.59% ±0.66	+2.4	91.80%	90.44% ±3.10	-1.4	52.44%	56.54% ±4.71	+4.1
28	e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	85.21%	87.25% ±1.35	+2.0	87.05%	89.32% ±0.83	+2.3	92.82%	90.00% ±4.07	-2.8	52.13%	55.90% ±6.84	+3.8	3 · 1942.0/1942.0	81.54%	83.71% ±2.31	+2.2	86.64%	89.03% ±0.93	+2.4	92.90%	89.89% ±4.21	-3.0	51.28%	55.15% ±7.16	+3.9
29	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1	m1	1 · 2358.0/2358.0	77.27%	79.25%	+2.0	87.00%	89.39%	+2.4	93.08%	92.31%	-0.8	53.07%	55.56%	+2.5	1 · 1942.0/1942.0	77.30%	79.41%	+2.1	87.05%	89.58%	+2.5	94.26%	93.44%	-0.8	54.25%	57.00%	+2.8
30	e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1	m1	1 · 2358.0/2358.0	86.48%	88.66%	+2.2	87.77%	90.39%	+2.6	96.15%	95.38%	-0.8	51.44%	56.88%	+5.4	1 · 1942.0/1942.0	83.88%	86.13%	+2.3	87.32%	90.03%	+2.7	95.90%	95.08%	-0.8	50.43%	55.77%	+5.3
31	e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1	m1	1 · 2358.0/2358.0	87.97%	90.16%	+2.2	89.85%	92.48%	+2.6	92.31%	92.31%	+0.0	61.86%	69.36%	+7.5	1 · 1942.0/1942.0	85.02%	87.30%	+2.3	89.52%	92.25%	+2.7	91.80%	91.80%	+0.0	60.87%	68.29%	+7.4
32	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_m1	m1	3 · 2358.0/2358.0	85.28%	86.56% ±0.31	+1.3	87.49%	88.71% ±0.73	+1.2	94.87%	92.56% ±1.18	-2.3	53.03%	54.47% ±1.03	+1.4	3 · 1942.0/1942.0	81.12%	83.16% ±0.83	+2.0	87.20%	88.53% ±0.61	+1.3	95.08%	92.62% ±0.82	-2.5	52.74%	54.25% ±1.16	+1.5
33	e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1	m1	3 · 2358.0/2358.0	85.09%	87.58% ±2.10	+2.5	87.87%	89.56% ±0.84	+1.7	93.33%	91.03% ±2.91	-2.3	54.46%	58.93% ±1.87	+4.5	3 · 1942.0/1942.0	80.48%	84.04% ±3.12	+3.6	87.52%	89.33% ±0.88	+1.8	93.44%	90.98% ±2.84	-2.5	53.74%	58.36% ±1.69	+4.6
34	e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	83.55%	86.21% ±0.36	+2.7	86.93%	89.31% ±1.13	+2.4	95.13%	93.85% ±2.31	-1.3	52.23%	56.28% ±6.08	+4.1	3 · 1942.0/1942.0	77.04%	79.76% ±0.74	+2.7	86.38%	88.89% ±1.18	+2.5	94.81%	93.44% ±2.46	-1.4	51.03%	55.16% ±6.23	+4.1
36	e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1	m1	3 · 2358.0/2358.0	87.26%	89.27% ±0.36	+2.0	89.57%	91.90% ±1.19	+2.3	94.36%	94.62% ±1.54	+0.3	60.20%	65.69% ±9.95	+5.5	3 · 1942.0/1942.0	83.97%	86.05% ±0.74	+2.1	89.25%	91.75% ±1.29	+2.5	93.99%	94.26% ±1.64	+0.3	59.47%	65.30% ±10.46	+5.8
37	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	84.91%	86.31% ±0.94	+1.4	87.18%	88.55% ±0.76	+1.4	92.05%	90.77% ±2.66	-1.3	51.47%	52.04% ±2.64	+0.6	3 · 1942.0/1942.0	79.93%	81.87% ±2.11	+1.9	86.81%	88.26% ±0.70	+1.5	91.80%	90.44% ±2.37	-1.4	50.77%	51.35% ±2.42	+0.6
38	e2_gemma3_1b_pt_cross_ce_real_lora_m1	m1	3 · 2358.0/2358.0	84.98%	87.29% ±1.71	+2.3	87.35%	89.15% ±0.91	+1.8	93.85%	91.28% ±1.60	-2.6	52.69%	56.09% ±2.91	+3.4	3 · 1942.0/1942.0	80.59%	83.58% ±2.71	+3.0	87.14%	89.04% ±0.91	+1.9	94.54%	91.80% ±2.17	-2.7	52.60%	56.19% ±2.57	+3.6
39	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_m1	m1	3 · 2358.0/2358.0	86.57%	88.56% ±0.49	+2.0	87.40%	89.12% ±0.38	+1.7	92.82%	92.31% ±0.77	-0.5	53.11%	54.34% ±1.97	+1.2	3 · 1942.0/1942.0	83.38%	85.79% ±0.89	+2.4	87.07%	88.89% ±0.50	+1.8	92.90%	92.35% ±0.95	-0.5	52.59%	53.79% ±2.27	+1.2
40	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m1	m1	3 · 2358.0/2358.0	86.33%	87.94% ±0.99	+1.6	88.16%	89.76% ±0.45	+1.6	91.54%	90.00% ±1.54	-1.5	58.20%	59.56% ±2.04	+1.4	3 · 1942.0/1942.0	82.44%	84.44% ±2.63	+2.0	87.90%	89.62% ±0.55	+1.7	91.80%	90.16% ±1.64	-1.6	57.72%	59.13% ±2.47	+1.4
41	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_m1	m1	3 · 2358.0/2358.0	83.85%	85.46% ±1.88	+1.6	86.93%	88.57% ±0.48	+1.6	93.08%	91.79% ±1.18	-1.3	50.22%	52.69% ±1.97	+2.5	3 · 1942.0/1942.0	79.38%	80.93% ±4.37	+1.5	86.56%	88.32% ±0.40	+1.8	93.17%	91.80% ±0.82	-1.4	49.43%	51.97% ±1.74	+2.5
42	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1	m1	3 · 2358.0/2358.0	83.60%	85.60% ±1.27	+2.0	86.67%	88.90% ±1.54	+2.2	91.79%	90.51% ±3.20	-1.3	53.54%	56.15% ±7.70	+2.6	3 · 1942.0/1942.0	79.32%	81.43% ±1.66	+2.1	86.22%	88.58% ±1.60	+2.4	91.80%	90.44% ±3.10	-1.4	52.68%	55.36% ±7.74	+2.7
43	e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	84.00%	86.30% ±0.90	+2.3	86.21%	88.30% ±0.27	+2.1	92.56%	91.28% ±2.70	-1.3	48.07%	52.66% ±1.35	+4.6	3 · 1942.0/1942.0	80.58%	83.17% ±1.42	+2.6	85.87%	88.10% ±0.36	+2.2	92.62%	91.26% ±2.50	-1.4	47.62%	52.44% ±1.67	+4.8
44	e2_qwen3_17b_base_cross_ce_real_lora_m1	m1	1 · 2358.0/2358.0	87.19%	88.76%	+1.6	88.91%	90.21%	+1.3	88.46%	88.46%	+0.0	56.65%	62.84%	+6.2	1 · 1942.0/1942.0	84.21%	86.12%	+1.9	88.56%	90.01%	+1.4	88.52%	88.52%	+0.0	55.96%	62.43%	+6.5
45	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m1	m1	3 · 2358.0/2358.0	85.40%	87.07% ±1.16	+1.7	87.52%	89.16% ±1.03	+1.6	93.33%	92.31% ±0.00	-1.0	52.32%	53.78% ±4.96	+1.5	3 · 1942.0/1942.0	80.57%	82.77% ±2.39	+2.2	87.13%	88.87% ±1.11	+1.7	93.17%	92.08% ±0.47	-1.1	51.65%	53.18% ±5.36	+1.5
46	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1	m1	3 · 2358.0/2358.0	86.68%	88.84% ±2.19	+2.2	88.27%	89.95% ±1.50	+1.7	94.10%	93.33% ±1.60	-0.8	55.55%	57.69% ±7.74	+2.1	3 · 1942.0/1942.0	81.99%	85.26% ±3.92	+3.3	87.99%	89.78% ±1.53	+1.8	94.26%	93.44% ±2.17	-0.8	55.17%	57.41% ±7.92	+2.2
47	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m1	m1	3 · 2358.0/2358.0	84.45%	86.51% ±1.76	+2.1	87.36%	89.14% ±0.89	+1.8	92.31%	92.05% ±0.89	-0.3	53.92%	54.95% ±3.19	+1.0	3 · 1942.0/1942.0	80.17%	82.44% ±3.24	+2.3	87.01%	88.89% ±0.95	+1.9	92.62%	92.08% ±0.47	-0.5	53.26%	54.34% ±3.42	+1.1
48	e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1	m1	3 · 2358.0/2358.0	84.71%	86.65% ±0.34	+1.9	86.71%	88.67% ±1.39	+2.0	93.08%	90.77% ±1.33	-2.3	50.62%	54.77% ±4.72	+4.2	3 · 1942.0/1942.0	80.82%	83.16% ±0.45	+2.3	86.40%	88.48% ±1.24	+2.1	93.72%	91.26% ±2.37	-2.5	50.05%	54.22% ±4.09	+4.2
49	e2_gemma3_270m_cross_ce_real_lora_m1	m1	3 · 2358.0/2358.0	82.03%	84.02% ±0.63	+2.0	85.10%	87.01% ±1.28	+1.9	90.77%	88.72% ±1.94	-2.1	48.54%	51.39% ±2.75	+2.8	3 · 1942.0/1942.0	78.01%	80.35% ±1.65	+2.3	84.77%	86.80% ±1.42	+2.0	91.26%	89.07% ±1.71	-2.2	48.16%	51.13% ±3.36	+3.0
50	e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1	m1	3 · 2358.0/2358.0	85.28%	87.12% ±0.75	+1.8	88.08%	89.89% ±0.51	+1.8	93.08%	90.77% ±1.33	-2.3	54.45%	58.77% ±1.88	+4.3	3 · 1942.0/1942.0	80.40%	82.43% ±1.51	+2.0	87.82%	89.73% ±0.45	+1.9	93.72%	91.26% ±1.25	-2.5	53.81%	58.27% ±2.38	+4.5
51	e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	85.14%	87.69% ±1.32	+2.6	86.29%	88.64% ±0.82	+2.3	94.10%	91.79% ±1.18	-2.3	47.86%	51.50% ±2.35	+3.6	3 · 1942.0/1942.0	82.83%	85.76% ±1.32	+2.9	86.44%	88.99% ±0.31	+2.5	93.99%	91.80% ±1.64	-2.2	49.72%	54.13% ±0.86	+4.4
52	e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	83.95%	86.24% ±2.47	+2.3	86.35%	88.39% ±0.56	+2.0	93.08%	91.03% ±1.94	-2.1	49.95%	53.98% ±1.11	+4.0	3 · 1942.0/1942.0	81.02%	83.55% ±3.43	+2.5	86.24%	88.43% ±0.70	+2.2	93.72%	91.53% ±2.06	-2.2	50.45%	54.84% ±0.71	+4.4
53	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1	m1	3 · 2358.0/2358.0	82.79%	85.12% ±1.71	+2.3	85.97%	88.09% ±2.33	+2.1	94.10%	92.82% ±1.18	-1.3	47.82%	51.02% ±5.20	+3.2	3 · 1942.0/1942.0	77.64%	80.09% ±1.19	+2.5	85.51%	87.75% ±2.24	+2.2	93.99%	92.62% ±0.82	-1.4	46.88%	50.15% ±4.86	+3.3
54	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1	m1	3 · 2358.0/2358.0	84.47%	85.95% ±1.42	+1.5	87.13%	89.23% ±0.41	+2.1	93.59%	92.05% ±1.94	-1.5	51.27%	55.09% ±3.53	+3.8	3 · 1942.0/1942.0	80.92%	82.46% ±1.90	+1.5	86.82%	89.04% ±0.36	+2.2	93.72%	92.08% ±1.71	-1.6	50.96%	54.96% ±2.11	+4.0
55	e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m1	m1	1 · 2358.0/2358.0	85.72%	88.21%	+2.5	87.69%	90.13%	+2.4	92.31%	90.00%	-2.3	56.60%	61.58%	+5.0	1 · 1942.0/1942.0	82.69%	85.39%	+2.7	87.46%	89.97%	+2.5	92.62%	90.16%	-2.5	56.50%	61.45%	+5.0
56	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	84.20%	86.43% ±1.26	+2.2	86.75%	88.87% ±0.60	+2.1	91.28%	90.77% ±1.33	-0.5	50.94%	54.69% ±2.02	+3.8	3 · 1942.0/1942.0	79.67%	82.07% ±2.53	+2.4	86.38%	88.63% ±0.64	+2.2	91.53%	90.98% ±1.42	-0.5	50.31%	54.23% ±2.32	+3.9
57	e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1	m1	3 · 2358.0/2358.0	87.34%	89.48% ±1.48	+2.1	88.53%	90.43% ±1.05	+1.9	91.28%	91.03% ±1.60	-0.3	54.11%	58.80% ±3.98	+4.7	3 · 1942.0/1942.0	84.40%	85.68% ±3.76	+1.3	88.07%	90.07% ±1.17	+2.0	90.98%	90.71% ±1.25	-0.3	52.95%	57.59% ±4.38	+4.6
58	e2_xlmr_base_khonly_ce_real_frozenemb_m1	m1	1 · 2358.0/2358.0	83.09%	85.23%	+2.1	86.38%	88.36%	+2.0	94.62%	93.08%	-1.5	50.62%	53.54%	+2.9	1 · 1942.0/1942.0	76.70%	78.93%	+2.2	85.89%	88.05%	+2.2	95.08%	93.44%	-1.6	49.57%	52.78%	+3.2
59	e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1	m1	3 · 2358.0/2358.0	85.08%	87.21% ±1.12	+2.1	86.24%	88.57% ±0.11	+2.3	93.85%	93.33% ±2.35	-0.5	49.59%	52.95% ±3.34	+3.4	3 · 1942.0/1942.0	81.44%	83.78% ±2.31	+2.3	86.00%	88.44% ±0.15	+2.4	94.26%	93.72% ±2.50	-0.5	49.53%	53.02% ±3.03	+3.5
60	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m1	m1	3 · 2358.0/2358.0	85.13%	87.13% ±1.03	+2.0	86.89%	88.69% ±0.49	+1.8	94.10%	93.33% ±0.44	-0.8	49.35%	51.14% ±0.87	+1.8	3 · 1942.0/1942.0	80.19%	82.90% ±2.30	+2.7	86.50%	88.37% ±0.49	+1.9	93.99%	93.17% ±0.47	-0.8	48.60%	50.30% ±0.56	+1.7
61	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m1	m1	3 · 2358.0/2358.0	85.03%	86.62% ±1.96	+1.6	86.90%	88.14% ±1.56	+1.2	92.05%	90.26% ±2.70	-1.8	51.77%	51.57% ±5.09	-0.2	3 · 1942.0/1942.0	80.26%	82.22% ±3.77	+2.0	86.54%	87.91% ±1.58	+1.4	92.35%	90.44% ±2.88	-1.9	50.93%	50.85% ±5.09	-0.1
62	e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m1	m1	3 · 2358.0/2358.0	85.50%	87.25% ±1.21	+1.8	87.20%	89.07% ±0.80	+1.9	94.10%	93.33% ±0.89	-0.8	53.44%	54.45% ±4.56	+1.0	3 · 1942.0/1942.0	81.04%	82.91% ±2.44	+1.9	86.84%	88.75% ±0.82	+1.9	94.26%	93.44% ±0.82	-0.8	52.69%	53.64% ±4.72	+1.0
63	e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1	m1	3 · 2358.0/2358.0	81.15%	83.17% ±4.68	+2.0	87.39%	89.02% ±1.32	+1.6	90.77%	88.21% ±0.44	-2.6	53.06%	56.43% ±5.94	+3.4	3 · 1942.0/1942.0	77.58%	80.31% ±2.96	+2.7	87.10%	88.87% ±1.39	+1.8	90.71%	88.25% ±0.47	-2.5	53.09%	56.67% ±6.46	+3.6
64	e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1	m1	3 · 2358.0/2358.0	83.72%	85.26% ±1.02	+1.5	86.09%	88.36% ±0.15	+2.3	93.85%	92.56% ±2.35	-1.3	47.34%	51.45% ±0.81	+4.1	3 · 1942.0/1942.0	79.65%	81.05% ±3.57	+1.4	85.99%	88.41% ±0.14	+2.4	94.26%	92.90% ±2.50	-1.4	47.88%	52.34% ±1.19	+4.5
65	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	84.72%	86.91% ±1.10	+2.2	87.64%	90.02% ±0.55	+2.4	92.56%	91.03% ±1.78	-1.5	53.52%	58.03% ±1.21	+4.5	3 · 1942.0/1942.0	77.00%	79.33% ±1.09	+2.3	87.35%	89.91% ±0.62	+2.6	93.17%	91.80% ±2.17	-1.4	53.00%	57.76% ±1.41	+4.8
66	e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1	m1	3 · 2358.0/2358.0	85.49%	88.02% ±2.03	+2.5	88.06%	90.33% ±1.33	+2.3	90.77%	89.49% ±0.89	-1.3	57.55%	61.24% ±7.00	+3.7	3 · 1942.0/1942.0	79.69%	82.79% ±4.25	+3.1	87.85%	90.27% ±1.27	+2.4	91.80%	90.44% ±0.95	-1.4	57.07%	61.00% ±6.57	+3.9
67	e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	84.81%	86.48% ±1.03	+1.7	87.17%	88.71% ±0.88	+1.5	92.56%	91.54% ±1.33	-1.0	52.98%	53.33% ±2.71	+0.3	3 · 1942.0/1942.0	80.53%	82.64% ±1.93	+2.1	86.78%	88.51% ±0.97	+1.7	92.62%	91.53% ±0.95	-1.1	52.41%	52.99% ±3.16	+0.6
68	e2_xlmr_base_cross_logitadjust_real_frozenemb_m1	m1	3 · 2358.0/2358.0	82.89%	85.63% ±0.67	+2.7	85.62%	88.15% ±0.92	+2.5	93.59%	92.56% ±0.89	-1.0	47.18%	50.52% ±2.90	+3.3	3 · 1942.0/1942.0	78.09%	81.09% ±0.77	+3.0	85.13%	87.82% ±1.06	+2.7	93.99%	92.90% ±1.71	-1.1	46.40%	49.83% ±3.25	+3.4
69	e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	86.24%	88.63% ±0.89	+2.4	88.85%	90.75% ±0.78	+1.9	92.05%	90.26% ±1.78	-1.8	58.24%	62.32% ±6.19	+4.1	3 · 1942.0/1942.0	82.41%	85.18% ±1.02	+2.8	88.45%	90.46% ±0.76	+2.0	91.80%	89.89% ±2.37	-1.9	57.20%	61.40% ±6.06	+4.2
70	e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1	m1	3 · 2358.0/2358.0	84.02%	85.79% ±1.12	+1.8	86.95%	88.94% ±0.35	+2.0	89.23%	87.44% ±4.70	-1.8	54.26%	58.10% ±3.91	+3.8	3 · 1942.0/1942.0	79.33%	81.12% ±2.25	+1.8	86.53%	88.63% ±0.40	+2.1	89.34%	87.43% ±5.01	-1.9	53.28%	57.23% ±3.66	+3.9
71	e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1	m1	3 · 2358.0/2358.0	83.85%	85.77% ±0.44	+1.9	86.97%	89.12% ±0.48	+2.2	93.08%	90.00% ±1.54	-3.1	53.05%	57.59% ±2.12	+4.5	3 · 1942.0/1942.0	79.37%	81.34% ±1.45	+2.0	86.58%	88.87% ±0.62	+2.3	93.44%	90.16% ±1.64	-3.3	52.36%	57.06% ±2.28	+4.7
72	e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1	m1	1 · 2358.0/2358.0	83.94%	85.75%	+1.8	85.91%	88.08%	+2.2	93.08%	93.08%	+0.0	44.98%	49.19%	+4.2	1 · 1942.0/1942.0	80.17%	82.11%	+1.9	85.54%	87.86%	+2.3	93.44%	93.44%	+0.0	44.53%	48.93%	+4.4
73	e2_visobert_cross_ce_real_frozenemb_lr4e5_m1	m1	1 · 2358.0/2358.0	81.50%	82.96%	+1.5	83.40%	85.15%	+1.7	93.08%	92.31%	-0.8	39.16%	40.54%	+1.4	1 · 1942.0/1942.0	77.56%	79.13%	+1.6	83.07%	84.95%	+1.9	93.44%	92.62%	-0.8	39.18%	40.65%	+1.5
74	e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1	m1	1 · 2358.0/2358.0	83.69%	86.87%	+3.2	86.02%	88.59%	+2.6	89.23%	89.23%	+0.0	47.35%	51.33%	+4.0	1 · 1942.0/1942.0	79.12%	82.79%	+3.7	85.72%	88.44%	+2.7	90.16%	90.16%	+0.0	47.21%	51.40%	+4.2
75	e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	85.18%	86.53% ±0.93	+1.3	87.47%	89.09% ±1.39	+1.6	93.59%	91.79% ±1.60	-1.8	50.41%	53.60% ±6.21	+3.2	3 · 1942.0/1942.0	79.32%	80.79% ±2.08	+1.5	87.22%	88.98% ±1.25	+1.8	93.72%	92.08% ±1.71	-1.6	50.15%	53.57% ±5.27	+3.4
76	e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	86.95%	88.27% ±0.46	+1.3	89.28%	90.71% ±1.97	+1.4	88.72%	87.95% ±3.47	-0.8	66.36%	65.70% ±12.82	-0.7	3 · 1942.0/1942.0	83.55%	85.19% ±1.18	+1.6	88.94%	90.43% ±2.10	+1.5	88.80%	87.70% ±3.28	-1.1	65.57%	64.74% ±13.13	-0.8
77	e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	85.76%	87.43% ±0.72	+1.7	88.15%	89.96% ±0.35	+1.8	90.51%	89.74% ±1.78	-0.8	57.48%	58.32% ±3.56	+0.8	3 · 1942.0/1942.0	80.88%	82.66% ±2.22	+1.8	87.70%	89.64% ±0.48	+1.9	90.71%	89.89% ±1.89	-0.8	56.77%	57.62% ±3.75	+0.9
78	e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_m1	m1	3 · 2358.0/2358.0	84.57%	86.72% ±0.33	+2.2	87.68%	89.31% ±1.65	+1.6	91.54%	89.74% ±3.87	-1.8	56.25%	57.40% ±8.00	+1.2	3 · 1942.0/1942.0	78.76%	81.96% ±1.18	+3.2	87.27%	89.02% ±1.64	+1.8	91.53%	89.62% ±4.21	-1.9	55.26%	56.52% ±7.79	+1.3
79	e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_m1	m1	3 · 2358.0/2358.0	84.59%	85.74% ±0.40	+1.2	87.37%	88.95% ±1.34	+1.6	92.56%	90.51% ±2.70	-2.1	55.37%	55.59% ±5.54	+0.2	3 · 1942.0/1942.0	78.83%	80.05% ±1.85	+1.2	87.09%	88.81% ±1.29	+1.7	92.62%	90.44% ±2.88	-2.2	55.01%	55.41% ±5.35	+0.4
80	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1	m1	3 · 2358.0/2358.0	83.85%	86.18% ±1.50	+2.3	86.80%	88.71% ±0.89	+1.9	91.54%	90.51% ±1.18	-1.0	51.50%	54.54% ±4.14	+3.0	3 · 1942.0/1942.0	78.96%	81.37% ±3.13	+2.4	86.33%	88.35% ±0.96	+2.0	91.53%	90.44% ±1.71	-1.1	50.29%	53.37% ±4.09	+3.1
81	e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	84.98%	86.95% ±0.66	+2.0	86.76%	88.87% ±1.06	+2.1	91.03%	89.23% ±4.62	-1.8	51.91%	54.54% ±5.99	+2.6	3 · 1942.0/1942.0	82.24%	84.13% ±1.55	+1.9	86.38%	88.60% ±1.05	+2.2	90.98%	89.07% ±4.51	-1.9	51.25%	53.98% ±5.72	+2.7
82	e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1	m1	3 · 2358.0/2358.0	85.37%	87.69% ±1.41	+2.3	87.93%	90.16% ±0.89	+2.2	91.79%	90.26% ±1.94	-1.5	57.88%	61.45% ±5.92	+3.6	3 · 1942.0/1942.0	81.35%	83.87% ±1.95	+2.5	87.62%	89.91% ±0.88	+2.3	92.08%	90.44% ±1.71	-1.6	57.03%	60.61% ±5.84	+3.6
83	e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1	m1	3 · 2358.0/2358.0	85.05%	87.36% ±1.75	+2.3	87.47%	89.58% ±2.23	+2.1	93.59%	93.59% ±1.18	-0.0	50.42%	55.62% ±8.24	+5.2	3 · 1942.0/1942.0	81.55%	84.41% ±2.42	+2.9	87.05%	89.33% ±2.35	+2.3	93.44%	93.44% ±1.42	+0.0	49.73%	55.22% ±8.43	+5.5
84	e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1	m1	3 · 2358.0/2358.0	84.16%	85.40% ±2.21	+1.2	85.86%	87.35% ±0.99	+1.5	90.26%	88.97% ±3.87	-1.3	45.47%	47.72% ±4.30	+2.3	3 · 1942.0/1942.0	80.31%	81.69% ±4.11	+1.4	85.90%	87.55% ±1.53	+1.7	90.44%	89.07% ±4.13	-1.4	47.40%	50.16% ±7.15	+2.8
85	e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_m1	m1	3 · 2358.0/2358.0	86.17%	87.99% ±0.91	+1.8	87.95%	89.57% ±0.44	+1.6	92.56%	91.28% ±1.78	-1.3	55.82%	56.78% ±0.15	+1.0	3 · 1942.0/1942.0	82.81%	84.89% ±1.45	+2.1	87.57%	89.26% ±0.52	+1.7	92.62%	91.26% ±2.37	-1.4	54.88%	55.85% ±0.16	+1.0
86	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m1	m1	3 · 2358.0/2358.0	86.35%	87.42% ±1.11	+1.1	88.02%	89.56% ±2.06	+1.5	91.54%	90.00% ±2.77	-1.5	58.93%	59.09% ±8.12	+0.2	3 · 1942.0/1942.0	82.19%	83.69% ±1.30	+1.5	87.64%	89.29% ±2.08	+1.6	90.98%	89.34% ±2.96	-1.6	58.33%	58.48% ±8.17	+0.2
87	e2_xlmr_base_cross_ce_real_frozenemb_m1	m1	3 · 2358.0/2358.0	83.20%	85.94% ±1.86	+2.7	86.03%	88.28% ±1.37	+2.2	90.26%	90.26% ±3.55	+0.0	50.47%	54.11% ±8.05	+3.6	3 · 1942.0/1942.0	77.52%	80.40% ±3.98	+2.9	85.71%	88.04% ±1.57	+2.3	90.98%	90.98% ±3.57	+0.0	49.96%	53.67% ±8.66	+3.7
88	e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1	m1	3 · 2358.0/2358.0	84.31%	86.94% ±0.72	+2.6	86.38%	88.96% ±1.02	+2.6	92.56%	91.54% ±2.66	-1.0	51.70%	55.65% ±6.10	+3.9	3 · 1942.0/1942.0	80.41%	83.32% ±1.04	+2.9	85.92%	88.68% ±1.16	+2.8	93.44%	92.35% ±1.71	-1.1	50.80%	54.92% ±6.34	+4.1
89	e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m1	m1	3 · 2358.0/2358.0	83.91%	84.70% ±1.42	+0.8	87.05%	88.27% ±0.47	+1.2	93.08%	91.54% ±0.77	-1.5	51.80%	51.60% ±0.80	-0.2	3 · 1942.0/1942.0	79.40%	80.28% ±2.06	+0.9	86.76%	88.06% ±0.47	+1.3	93.17%	91.26% ±1.25	-1.9	51.38%	51.16% ±0.84	-0.2
90	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m1	m1	3 · 2358.0/2358.0	83.15%	84.00% ±4.29	+0.9	87.48%	88.88% ±0.88	+1.4	91.03%	90.00% ±2.04	-1.0	56.24%	55.57% ±7.22	-0.7	3 · 1942.0/1942.0	80.71%	81.30% ±2.29	+0.6	87.22%	88.66% ±0.85	+1.4	91.53%	89.34% ±2.17	-2.2	55.95%	55.16% ±6.84	-0.8
91	e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1	m1	2 · 2358.0/2358.0	82.11%	83.59% ±0.67	+1.5	84.23%	85.84% ±1.88	+1.6	90.00%	88.85% ±2.72	-1.2	47.85%	50.88% ±0.03	+3.0	2 · 1942.0/1942.0	78.03%	79.60% ±0.27	+1.6	83.47%	85.17% ±2.01	+1.7	89.34%	88.11% ±2.90	-1.2	46.58%	49.53% ±0.34	+3.0
92	e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1	m1	1 · 2358.0/2358.0	76.27%	76.73%	+0.5	86.53%	89.22%	+2.7	89.23%	88.46%	-0.8	54.98%	57.50%	+2.5	1 · 1942.0/1942.0	75.77%	74.44%	-1.3	86.47%	89.32%	+2.8	90.16%	89.34%	-0.8	55.84%	58.60%	+2.8
93	e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1	m1	3 · 2358.0/2358.0	82.32%	84.55% ±0.52	+2.2	84.83%	86.97% ±1.13	+2.1	84.10%	82.82% ±0.44	-1.3	52.38%	54.84% ±5.54	+2.5	3 · 1942.0/1942.0	78.33%	80.80% ±0.59	+2.5	84.15%	86.42% ±1.26	+2.3	83.33%	81.97% ±0.00	-1.4	51.01%	53.51% ±5.90	+2.5
94	e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m1	m1	3 · 2358.0/2358.0	85.20%	86.62% ±0.35	+1.4	87.48%	88.72% ±0.95	+1.2	93.59%	92.05% ±1.94	-1.5	53.58%	54.09% ±5.37	+0.5	3 · 1942.0/1942.0	80.89%	82.56% ±0.15	+1.7	87.07%	88.41% ±0.98	+1.3	93.72%	92.08% ±2.06	-1.6	52.68%	53.18% ±5.32	+0.5
95	e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1	m1	1 · 2358.0/2358.0	76.82%	76.85%	+0.0	85.51%	87.51%	+2.0	90.00%	89.23%	-0.8	47.95%	50.88%	+2.9	1 · 1942.0/1942.0	77.06%	76.10%	-1.0	85.20%	87.32%	+2.1	90.16%	89.34%	-0.8	48.03%	51.17%	+3.1
96	e2_visobert_cross_ce_real_frozenemb_lr2e5_m1	m1	1 · 2358.0/2358.0	79.68%	82.81%	+3.1	83.95%	85.46%	+1.5	90.00%	89.23%	-0.8	41.79%	43.28%	+1.5	1 · 1942.0/1942.0	72.43%	77.08%	+4.6	83.58%	85.22%	+1.6	90.98%	90.16%	-0.8	41.57%	43.14%	+1.6
97	e2_visobert_cross_ce_real_frozenemb_lr1e5_m1	m1	1 · 2358.0/2358.0	79.30%	82.07%	+2.8	82.85%	83.94%	+1.1	86.15%	87.69%	+1.5	44.09%	45.42%	+1.3	1 · 1942.0/1942.0	73.36%	76.25%	+2.9	82.31%	83.50%	+1.2	86.07%	87.70%	+1.6	43.57%	44.96%	+1.4
98	e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1	m1	1 · 2358.0/2358.0	80.10%	84.03%	+3.9	85.75%	87.50%	+1.8	90.00%	90.00%	+0.0	52.70%	54.42%	+1.7	1 · 1942.0/1942.0	73.36%	79.91%	+6.6	85.37%	87.32%	+2.0	90.98%	90.98%	+0.0	52.11%	53.88%	+1.8
99	e1_tfidf_hybrid_qwen3emb_lr_m1	m1	1 · 2358.0/2358.0	81.01%	82.08%	+1.1	83.30%	85.16%	+1.9	84.62%	83.85%	-0.8	46.81%	49.77%	+3.0	1 · 1942.0/1942.0	75.09%	74.39%	-0.7	82.84%	84.82%	+2.0	84.43%	83.61%	-0.8	46.61%	49.76%	+3.1
100	e1_tfidf_svmcal_ce_synth1x_fbmatched_m1	m1	1 · 2358.0/2358.0	83.64%	84.20%	+0.6	85.82%	87.13%	+1.3	81.54%	79.23%	-2.3	57.61%	59.54%	+1.9	1 · 1942.0/1942.0	77.77%	78.33%	+0.6	85.33%	86.72%	+1.4	81.15%	78.69%	-2.5	56.90%	58.90%	+2.0
101	e1_tfidf_svmcal_ce_synth1x_nofb_m1	m1	1 · 2358.0/2358.0	82.76%	83.43%	+0.7	85.40%	86.78%	+1.4	80.77%	78.46%	-2.3	54.97%	57.63%	+2.7	1 · 1942.0/1942.0	76.80%	77.51%	+0.7	84.89%	86.35%	+1.5	80.33%	77.87%	-2.5	54.14%	56.89%	+2.7
102	e1_tfidf_svmcal_ce_synth1x_noqparticle_m1	m1	1 · 2358.0/2358.0	83.27%	83.13%	-0.1	83.92%	85.35%	+1.4	84.62%	83.85%	-0.8	46.81%	48.88%	+2.1	1 · 1942.0/1942.0	79.82%	78.52%	-1.3	83.48%	85.00%	+1.5	84.43%	83.61%	-0.8	46.61%	48.80%	+2.2
104	e1_tfidf_svmcal_ce_synth1x_nofiller_m1	m1	1 · 2358.0/2358.0	82.99%	84.50%	+1.5	83.93%	85.74%	+1.8	87.69%	86.92%	-0.8	46.91%	49.34%	+2.4	1 · 1942.0/1942.0	78.49%	80.08%	+1.6	83.28%	85.19%	+1.9	87.70%	86.89%	-0.8	45.73%	48.18%	+2.5
105	e1_tfidf_svmcal_ce_synth1x_noofftopic_m1	m1	1 · 2358.0/2358.0	83.02%	82.92%	-0.1	84.23%	85.65%	+1.4	84.62%	83.85%	-0.8	48.25%	50.23%	+2.0	1 · 1942.0/1942.0	79.44%	78.32%	-1.1	83.90%	85.42%	+1.5	84.43%	83.61%	-0.8	48.58%	50.75%	+2.2
106	e2_dual_qwen3emb06b_concat_ce_real_frozen_m1	m1	3 · 2358.0/2358.0	80.49%	82.22% ±1.31	+1.7	81.14%	83.74% ±1.22	+2.6	85.90%	86.15% ±0.00	+0.3	45.40%	48.94% ±4.87	+3.5	3 · 1942.0/1942.0	74.43%	76.77% ±1.64	+2.3	80.44%	83.16% ±1.21	+2.7	85.52%	85.25% ±0.00	-0.3	44.57%	48.08% ±4.71	+3.5
107	e1_tfidf_svmcal_ce_synth1x_base_m1	m1	1 · 2358.0/2358.0	82.56%	82.62%	+0.1	83.68%	85.30%	+1.6	83.85%	83.08%	-0.8	46.58%	48.65%	+2.1	1 · 1942.0/1942.0	78.84%	77.89%	-1.0	83.18%	84.89%	+1.7	83.61%	82.79%	-0.8	46.15%	48.33%	+2.2
108	e1_tfidf_svmcal_ce_synth1x_matched_m1	m1	1 · 2358.0/2358.0	82.76%	83.20%	+0.4	85.40%	86.51%	+1.1	83.08%	80.77%	-2.3	55.10%	56.76%	+1.7	1 · 1942.0/1942.0	76.78%	77.26%	+0.5	84.86%	86.04%	+1.2	82.79%	80.33%	-2.5	54.30%	56.00%	+1.7
109	e1_tfidf_svmcal_ce_synth1x_nomoney_m1	m1	1 · 2358.0/2358.0	83.74%	83.26%	-0.5	84.49%	85.51%	+1.0	85.38%	83.85%	-1.5	48.68%	50.46%	+1.8	1 · 1942.0/1942.0	80.24%	78.59%	-1.6	83.98%	85.08%	+1.1	85.25%	83.61%	-1.6	48.15%	50.00%	+1.9
110	e1_tfidf_svmcal_ce_synth1x_fbonly_m1	m1	1 · 2358.0/2358.0	82.32%	81.80%	-0.5	83.40%	84.32%	+0.9	85.38%	83.85%	-1.5	43.87%	43.78%	-0.1	1 · 1942.0/1942.0	78.78%	77.22%	-1.6	83.11%	84.10%	+1.0	85.25%	83.61%	-1.6	44.64%	44.54%	-0.1

Phân tích theo mức ảnh hưởng

Mức	Method	Run	Kết luận
high	e2_xlmr_base_cross_ce_real_frozenemb	e2_xlmr_base_cross_ce_real_frozenemb_seed13_0a1472e9	On dev, xlm-roberta-base cross-encoder with frozen embeddings beats every TF-IDF baseline on the same data/format/decode: fallback recall 0.975 +- 0.002 vs 0.943 (svmcal), precision 0.893 +- 0.030 vs 0.852, missed fallbacks 6-7 vs 15, fallb

Cảnh báo registry

e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1: mixed code_commit across seeds: ['7f7b3058ec98d363388230b6c033902a5a2618ea', 'a014dd3f703d0c4cec6a40aca79c4dae6b52761d']
e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m1: mixed code_commit across seeds: ['7f7b3058ec98d363388230b6c033902a5a2618ea', '8938eacfdbc03f60d19177775a404dd82773feda', 'a014dd3f703d0c4cec6a40aca79c4dae6b52761d']
e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepeval_m1: mixed per_device_batch across seeds: ['16', '32']
e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m1: mixed code_commit across seeds: ['50d8c1f26346b6b89c3ce75b8905966ad88d9114', '7c34e09071ac90ecc10cdf634ee16e7528dd83f0']
e2_me5_base_cross_ce_real_frozenemb_lr2e5: mixed code_commit across seeds: ['1503a547d6e50972632609b1efc77487b25c445e', '4166114f070e9c23ccc502896964be5ca1b5900d']
e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1: mixed code_commit across seeds: ['7c34e09071ac90ecc10cdf634ee16e7528dd83f0', '9cf9db5b286edb99e4de1ae900642d1f923c7f0c']
e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1: mixed code_commit across seeds: ['7c34e09071ac90ecc10cdf634ee16e7528dd83f0', '7f7b3058ec98d363388230b6c033902a5a2618ea', '9cf9db5b286edb99e4de1ae900642d1f923c7f0c']
e2_mmbert_base_cross_ce_real_frozenemb_lr1e5: mixed code_commit across seeds: ['4166114f070e9c23ccc502896964be5ca1b5900d', '434035beb22304e80dea2164609290ba0f7ea331']
e2_mmbert_base_cross_ce_real_frozenemb_lr2e5: mixed code_commit across seeds: ['4166114f070e9c23ccc502896964be5ca1b5900d', '434035beb22304e80dea2164609290ba0f7ea331']
e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1: mixed code_commit across seeds: ['50d8c1f26346b6b89c3ce75b8905966ad88d9114', 'e65a5db21922265e4eed2d878504fec163e38d39']
e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5: mixed code_commit across seeds: ['1503a547d6e50972632609b1efc77487b25c445e', '4166114f070e9c23ccc502896964be5ca1b5900d']
e2_qwen3_17b_base_cross_ce_real_lora_m1: mixed code_commit across seeds: ['2d84d036fdb33cea12d8694b108da7192b7fbbf2', 'bd04004542a18993c9547205ccdae4d3ea09a0d8']
e2_xlmr_base_cross_ce_real_frozenemb: mixed code_commit across seeds: ['f9b4c786240ee5e4b48b0be346fa65edf96b8a8d', 'fa6989fead7d8ef372895c2cfbbdbd7a68a99369']