HappyCall · kết quả thí nghiệm   tiến độ

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 2026-10-26 17:44 UTC.

zephyrus · 260/260 đã chạy xong

m1: 205/205

v1: 49/49

v2: 1/1

v2_1: 5/5

queued 0 · running 0 · treo 0 · ok 253 · failed 7

legion · 78/79 đã chạy xong · 17 bị thay thế (code cũ)

m1: 74/75

v1: 4/4

queued 0 · running 1 · treo 0 · ok 77 · failed 1 · ETA ~26m

đang chạy: k2_e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_oof5f4_m1 seed 13 trên legion, bắt đầu 2026-10-26 15:23 UTC (đã 2h 21m)

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.

NhãnTênĐịnh nghĩaVí dụ (bot → khách)
omi_intent_repeatYêu cầu nhắc lạiKhách 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.”
callbackHẹn gọi lạiKhách 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.”
relationshipNgười thân, không phải chủ hợp đồngNgười 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_fallbackKhông hiểu / lạc đềCâu 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?”
negatePhủ nhậnKhách 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.”
affirmĐồng ýKhông 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)

Các split

Cách đọc cột

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 2026-10-26 17:44 UTC.

STTModel_NameBackbone Threshold DataHọMáySeedn_effParamsLatency ms clean Accuracyclean Macro_Preclean Macro_Recclean Macro_F1clean Macro_F1 −repclean Weighted_F1clean FB recallclean FB precisionclean argmax Macro_F1clean argmax Macro_F1 −repclean argmax FB recallclean argmax FB precisionreply-unseen Accuracyreply-unseen Macro_Prereply-unseen Macro_Recreply-unseen Macro_F1reply-unseen Macro_F1 −repreply-unseen Weighted_F1reply-unseen FB recallreply-unseen FB precisionreply-unseen argmax Macro_F1reply-unseen argmax Macro_F1 −repreply-unseen argmax FB recallreply-unseen argmax FB precisionclean fb→negateclean fb→affirmclean false fbTrạng tháiPhân tích
0 e0_prod_sft_hard_none_qc_v16_m1 tham chiếu · provenance: unknown Qwen-2B-Chat (SFT generative)N/A (CausalLM) m1referencezephyrus 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) v1referencezephyrus 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) v2referencezephyrus 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_1referencezephyrus 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_m2 –0.051 m2ensemblezephyrus 1 1 N/A 271.8 ms 94.42%88.12%91.33%88.68%90.76%94.90%99.20%56.36%91.26%94.34%87.20%87.20%93.56%84.34%87.16%84.58%90.59%94.13%100.00%55.71%88.52%94.46%88.03%87.29%0.00%0.80%4.39%okp85_not_met –
2 e3_ensemble_devtop1_familygeomean_tempdev_m2 –0.051 m2ensemblezephyrus 1 1 N/A 271.8 ms 94.72%88.43%91.61%89.07%91.24%95.14%99.20%58.22%91.26%94.34%87.20%87.20%93.88%84.62%87.45%84.96%91.04%94.38%100.00%57.35%88.52%94.46%88.03%87.29%0.00%0.80%4.07%okp85_not_met –
3 e3_ensemble_devtop1_familygeomean_temphard_m2 –0.051 m2ensemblezephyrus 1 1 N/A 271.8 ms 94.72%88.43%91.61%89.07%91.24%95.14%99.20%58.22%91.26%94.34%87.20%87.20%93.88%84.62%87.45%84.96%91.04%94.38%100.00%57.35%88.52%94.46%88.03%87.29%0.00%0.80%4.07%okp85_not_met –
4 e3_ensemble_devtop1_familymean_m2 –0.112 m2ensemblezephyrus 1 1 N/A 271.8 ms 94.76%88.82%93.04%90.10%91.45%95.14%98.40%59.42%90.98%94.00%84.80%86.89%93.93%85.86%90.26%87.16%91.26%94.39%99.15%58.59%88.22%94.10%85.47%86.96%0.00%1.60%3.84%okp85_not_met –
5 e3_ensemble_sweep_k2_m2 –0.110 m2ensemblezephyrus 1 1 N/A 106.3 ms 95.76%90.29%92.70%91.15%92.71%95.89%96.00%71.01%92.28%94.07%85.60%87.70%95.14%87.31%89.99%88.27%92.59%95.31%96.58%70.19%89.58%94.16%86.32%87.83%0.80%2.40%2.24%okp85_not_met –
6 e3_ensemble_sweep_k3_m2 –0.051 m2ensemblezephyrus 1 1 N/A 45.3 ms 95.11%86.80%91.30%88.39%91.67%95.39%96.80%63.02%89.71%93.37%85.60%82.31%94.35%82.19%87.15%83.92%91.48%94.71%97.44%61.96%86.19%93.43%86.32%82.11%0.00%1.60%3.25%okp85_not_met –
7 e3_ensemble_sweep_k4_m2 –0.058 m2ensemblezephyrus 1 1 N/A 94.7 ms 95.24%86.99%91.49%88.59%91.91%95.52%96.80%63.02%89.89%93.58%84.80%82.81%94.51%82.39%87.36%84.13%91.73%94.87%97.44%61.96%86.38%93.65%85.47%82.64%0.00%1.60%3.25%okp85_not_met –
8 e3_ensemble_sweep_k5_m2 –0.111 m2ensemblezephyrus 1 1 N/A 277.3 ms 96.19%91.80%93.49%92.25%93.30%96.33%97.60%71.35%92.43%94.25%86.40%87.10%95.67%89.81%90.82%89.81%93.23%95.83%98.29%70.55%89.75%94.36%87.18%87.18%0.00%1.60%2.24%okp85_not_met –
9 e3_ensemble_sweep_k6_m2 –0.087 m2ensemblezephyrus 1 1 N/A 277.3 ms 95.89%91.24%93.32%91.77%92.74%96.08%97.60%67.78%92.43%94.25%86.40%87.10%95.30%89.29%90.63%89.34%92.67%95.52%98.29%67.25%89.75%94.36%87.18%87.18%0.00%1.60%2.65%okp85_not_met –
10 e3_ensemble_sweep_k7_m2 –0.087 m2ensemblezephyrus 1 1 N/A 277.3 ms 95.89%91.24%93.32%91.77%92.74%96.08%97.60%67.78%92.43%94.25%86.40%87.10%95.30%89.29%90.63%89.34%92.67%95.52%98.29%67.25%89.75%94.36%87.18%87.18%0.00%1.60%2.65%okp85_not_met –
11 e3_ensemble_sweep_k8_m2 –0.087 m2ensemblezephyrus 1 1 N/A 277.3 ms 95.89%91.24%93.32%91.77%92.74%96.08%97.60%67.78%92.43%94.25%86.40%87.10%95.30%89.29%90.63%89.34%92.67%95.52%98.29%67.25%89.75%94.36%87.18%87.18%0.00%1.60%2.65%okp85_not_met –
12 e3_ensemble_top_geomean_m2 –0.073 m2ensemblezephyrus 1 1 N/A 149.4 ms 95.15%87.23%93.36%89.56%92.08%95.46%97.60%62.56%90.56%94.47%87.20%86.51%94.40%83.45%90.61%86.10%91.89%94.80%98.29%61.50%87.60%94.59%88.03%86.55%0.00%1.60%3.34%okp85_not_met –
13 e3_ensemble_top_mean_m2 –0.073 m2ensemblezephyrus 1 1 N/A 149.4 ms 93.81%86.01%92.84%88.18%90.43%94.43%98.40%53.02%90.36%94.24%84.80%86.89%92.87%82.32%90.04%84.74%90.26%93.61%99.15%52.49%87.39%94.35%85.47%86.96%0.00%0.80%4.99%okp85_not_met –
14 e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m2 unsloth/gemma-3-1b-pt0.064 m2decoderlegion 3 3 24M 94.0 ms 93.15% ±1.4987.53% ±0.6388.56% ±0.3786.69% ±0.8389.06% ±2.1093.76% ±1.1096.00% ±0.8053.41% ±8.3090.25% ±0.3092.42% ±0.4881.60% ±3.4983.56% ±3.3492.06% ±1.6882.75% ±1.3382.89% ±1.4980.95% ±1.6988.81% ±2.1392.76% ±1.2396.58% ±0.8552.82% ±8.1386.79% ±0.5692.31% ±0.5381.48% ±3.9583.57% ±3.302.13%1.60%4.97%cần sửap85_not_met –
15 e2_me5_base_cross_ce_real_frozenemb_lr4e5_m2 intfloat/multilingual-e5-base0.017 m2encoderzephyrus 1 1 86M 11.7 ms 92.03%81.03%89.32%83.62%87.49%92.80%96.00%49.38%87.96%91.81%83.20%81.25%90.55%76.31%84.75%78.71%86.95%91.50%95.73%48.07%84.76%91.71%83.76%80.99%0.80%0.80%5.63%okp85_not_met –
16 e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m2 jhu-clsp/mmBERT-base0.155 m2encoderzephyrus 1 1 111M 22.4 ms 94.24%86.23%89.54%87.41%89.51%94.44%88.00%63.58%87.79%91.15%80.00%76.34%93.24%82.30%86.63%83.83%89.17%93.51%88.03%62.42%84.51%90.89%79.49%75.61%2.40%8.00%2.88%okp85_not_met –
17 e2_qwen3_17b_base_genlabel_ce_real_lora_m2 Qwen/Qwen3-1.7B-Base0.081 m2decoderlegion 1 1 17M 55.4 ms 93.85%84.73%92.99%87.74%89.57%94.35%93.60%55.45%87.62%91.39%79.20%74.44%92.77%81.31%91.35%84.74%89.19%93.39%93.16%54.23%84.38%91.25%79.49%73.81%1.60%2.40%4.30%okp85_not_met –
18 e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m2 Qwen/Qwen3-1.7B-Base0.105 m2decoderlegion 3 3 17M 52.8 ms 93.98% ±0.2087.30% ±0.8293.59% ±0.4189.45% ±0.2189.96% ±0.2694.50% ±0.2095.47% ±0.4654.82% ±0.0389.75% ±1.0492.64% ±0.3485.60% ±1.3977.25% ±5.1392.94% ±0.2284.70% ±1.0591.91% ±0.3387.17% ±0.4289.59% ±0.2493.57% ±0.2295.16% ±0.4953.87% ±0.2686.81% ±1.3092.50% ±0.3685.75% ±1.7876.88% ±5.071.87%1.33%4.50%okp85_not_met –
19 e3_ensemble_top_greedy_m2 –0.113 m2ensemblezephyrus 1 1 N/A 149.4 ms 94.59%86.93%94.22%89.56%91.18%95.02%96.80%57.89%91.59%94.73%87.20%89.34%93.72%83.78%92.80%86.91%90.96%94.25%97.44%57.00%88.87%94.88%88.03%89.57%0.80%1.60%4.03%okp85_not_met –
20 e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m2 intfloat/multilingual-e5-large-instruct0.017 m2encoderlegion 3 3 304M 12.3 ms 94.13% ±0.8986.17% ±0.6091.58% ±1.0087.94% ±0.3390.26% ±1.6794.53% ±0.6896.27% ±0.4659.78% ±8.7088.30% ±1.0992.38% ±0.2185.07% ±0.9279.78% ±1.8093.17% ±1.0182.49% ±0.8488.70% ±1.0984.46% ±0.0390.01% ±1.7193.67% ±0.7796.58% ±0.8558.92% ±8.5984.64% ±1.1692.34% ±0.2685.75% ±1.3179.45% ±1.730.53%1.87%3.83%okp85_not_met –
21 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m2 jhu-clsp/mmBERT-base0.094 m2encoderzephyrus 3 3 111M 20.7 ms 93.26% ±1.4285.68% ±1.6089.22% ±3.2186.25% ±2.9588.98% ±1.9793.83% ±1.1295.73% ±3.2354.90% ±9.3789.13% ±0.1991.88% ±0.5880.53% ±2.0181.18% ±0.7592.12% ±1.7180.15% ±3.7884.44% ±6.1980.85% ±5.5288.68% ±2.1192.78% ±1.3696.01% ±3.2454.16% ±9.7886.12% ±0.2891.91% ±0.5681.48% ±1.7881.24% ±1.020.53%2.67%4.71%okp85_not_met –
22 e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m2 Qwen/Qwen3-4B-Base0.056 m2decoderlegion 3 3 33M 138.1 ms 94.29% ±0.3485.74% ±0.6592.06% ±0.6687.97% ±0.3990.69% ±0.5194.71% ±0.2896.27% ±0.4657.52% ±1.6590.01% ±0.2793.43% ±0.8482.13% ±3.0384.50% ±3.7393.38% ±0.3881.66% ±0.5688.54% ±1.3484.02% ±0.6790.38% ±0.6093.89% ±0.3196.01% ±0.4956.86% ±1.5486.64% ±0.1693.38% ±0.9782.34% ±2.7584.90% ±4.361.07%1.33%4.07%okp85_not_met –
23 e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m2 unsloth/gemma-3-1b-pt0.045 m2decoderlegion 2 2 13M 51.1 ms 93.40% ±0.2886.52% ±2.4089.97% ±1.3487.08% ±0.3589.12% ±0.9793.98% ±0.2296.40% ±0.5753.23% ±1.8189.23% ±0.2292.51% ±0.8986.80% ±2.8378.90% ±0.5492.32% ±0.3482.86% ±3.7285.70% ±1.3782.88% ±0.8088.84% ±1.0093.01% ±0.2897.01% ±0.6052.45% ±1.8786.00% ±0.0792.49% ±0.9287.61% ±3.0278.54% ±0.582.00%1.60%4.85%okp85_not_met –
24 e2_gemma3_1b_pt_cross_ce_real_lora_r32_m2 unsloth/gemma-3-1b-pt0.045 m2decoderlegion 3 3 26M 49.5 ms 92.66% ±1.1284.56% ±1.9890.28% ±1.9686.15% ±2.1188.26% ±1.8693.33% ±0.8495.47% ±0.9251.30% ±6.0587.35% ±0.5491.63% ±0.2181.33% ±2.4480.08% ±0.9991.45% ±1.3780.73% ±2.8486.87% ±2.8582.34% ±2.9287.95% ±1.9792.26% ±1.0496.01% ±0.9950.48% ±6.2483.79% ±0.7791.60% ±0.2281.77% ±1.7879.98% ±1.431.87%2.13%5.28%okp85_not_met –
25 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m2 jhu-clsp/mmBERT-base0.052 m2encoderzephyrus 3 3 111M 18.0 ms 92.97% ±0.6482.86% ±1.0791.51% ±1.1685.65% ±1.0788.22% ±0.7393.51% ±0.5092.80% ±1.3954.21% ±4.2387.63% ±0.4991.71% ±0.7082.93% ±4.6279.58% ±3.6892.05% ±0.6379.72% ±1.1889.97% ±2.4082.70% ±1.4488.10% ±0.7192.66% ±0.5192.88% ±1.3154.50% ±3.5084.60% ±1.0491.60% ±0.7883.19% ±4.7179.18% ±3.941.33%4.27%4.53%okp85_not_met –
26 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m2 vinai/phobert-base-v20.008 m2encoderzephyrus 1 1 86M 12.0 ms 93.68%89.53%80.34%79.10%89.21%93.99%95.20%56.13%88.60%92.13%83.20%84.55%93.08%89.68%80.46%79.25%89.39%93.47%96.58%57.36%85.62%92.21%84.62%84.62%2.40%1.60%4.26%okp85_not_met –
27 e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m2 Qwen/Qwen3-4B-Base0.162 m2decoderkaggle 1 1 33M 233.2 ms 93.64%85.03%95.05%88.56%89.73%94.13%97.60%54.95%90.09%92.11%84.80%75.71%92.61%82.13%94.83%86.26%89.39%93.21%97.44%54.03%87.71%91.92%84.62%75.00%0.00%0.00%4.58%okp85_not_met –
28 e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m2 Qwen/Qwen3-8B-Base0.170 m2decoderkaggle 1 1 44M 379.1 ms 95.24%86.94%94.04%89.89%91.57%95.45%94.40%66.29%89.83%93.13%84.80%83.46%94.51%83.77%92.56%87.25%91.36%94.78%94.02%65.48%86.71%92.95%83.76%83.05%2.40%1.60%2.75%okp85_not_met –
29 e3_ensemble_seed6_kd_a05_t1_m2 –0.067 m2ensemblelegion 1 1 N/A 19.6 ms 94.59%89.40%91.35%89.54%91.09%94.91%96.80%60.80%92.27%93.80%85.60%86.29%93.77%86.74%87.20%85.79%90.95%94.13%97.44%60.32%90.03%93.75%85.47%86.21%0.00%2.40%3.57%okp85_not_met –
30 e2_me5_base_cross_ce_real_frozenemb_lr2e5_m2 intfloat/multilingual-e5-base0.031 m2encoderzephyrus 3 3 86M 9.0 ms 93.26% ±0.9384.64% ±0.8188.47% ±0.4585.51% ±0.6889.10% ±1.4293.77% ±0.7396.53% ±1.8555.47% ±5.8286.44% ±0.7791.03% ±1.0483.20% ±2.7773.93% ±7.3492.06% ±1.0778.19% ±0.5882.18% ±1.1179.06% ±0.6088.69% ±1.4792.66% ±0.8296.30% ±1.9754.39% ±5.9182.15% ±1.2591.03% ±1.0384.90% ±3.4573.61% ±7.461.33%1.33%4.52%okp85_not_met –
31 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m2 jhu-clsp/mmBERT-base0.061 m2encoderzephyrus 3 3 111M 17.6 ms 92.97% ±0.9284.49% ±1.1989.84% ±0.6986.01% ±0.9188.64% ±1.2293.55% ±0.6895.47% ±1.6753.62% ±6.4587.96% ±1.4891.91% ±0.9481.60% ±2.8880.48% ±5.4991.92% ±1.0680.24% ±1.4785.98% ±1.0981.84% ±1.0588.40% ±1.2892.60% ±0.8095.73% ±1.7153.19% ±6.0284.57% ±1.8491.80% ±1.0581.77% ±2.7580.13% ±5.870.80%2.40%4.84%okp85_not_met –
32 e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m2 Qwen/Qwen3-4B-Base0.131 m2decoderkaggle 3 3 33M 232.8 ms 94.95% ±0.8487.65% ±2.6893.06% ±1.0489.41% ±0.8691.57% ±1.2995.27% ±0.6595.73% ±2.0163.85% ±10.4090.54% ±0.7593.74% ±0.5184.53% ±2.5785.43% ±1.0394.16% ±1.0185.15% ±4.6690.62% ±1.9086.46% ±1.7191.35% ±1.3594.55% ±0.8095.44% ±2.1563.18% ±10.6587.51% ±0.8893.78% ±0.4684.62% ±2.2685.83% ±0.761.87%0.53%3.26%okp85_not_met –
33 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.043 m2encoderzephyrus 3 3 86M 20.9 ms 93.29% ±0.4986.97% ±2.8889.52% ±1.9987.11% ±1.8789.20% ±0.7493.80% ±0.4094.13% ±2.5754.24% ±2.2289.29% ±1.0091.86% ±0.2586.40% ±0.8073.16% ±1.7592.22% ±0.5283.95% ±5.6784.73% ±3.6482.44% ±3.2788.93% ±0.7192.81% ±0.4494.02% ±2.2653.56% ±2.0685.91% ±1.9491.65% ±0.2886.04% ±0.4972.62% ±1.903.20%1.60%4.56%okp85_not_met –
34 e2_gemma3_1b_pt_cross_ce_real_lora_m2 unsloth/gemma-3-1b-pt0.053 m2decoderlegion 3 3 13M 49.8 ms 93.48% ±0.7485.45% ±1.3989.55% ±0.8086.52% ±1.0189.19% ±1.1093.96% ±0.6695.20% ±0.0055.38% ±3.1188.74% ±2.2192.32% ±0.6983.73% ±2.0182.04% ±1.9092.52% ±0.8281.16% ±1.7485.37% ±0.7982.19% ±1.0789.06% ±1.0593.08% ±0.7496.01% ±0.4955.26% ±2.8685.41% ±3.1992.30% ±0.7384.33% ±2.1582.04% ±1.642.40%1.87%4.41%okp85_not_met –
35 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m2 vinai/phobert-base-v20.085 m2encoderzephyrus 3 3 86M 20.7 ms 94.24% ±0.3586.86% ±0.8291.31% ±2.1388.44% ±1.1790.29% ±0.3794.53% ±0.3293.60% ±0.8061.33% ±2.2789.66% ±0.5992.52% ±0.4084.00% ±1.6082.68% ±1.3793.40% ±0.5483.10% ±2.1788.02% ±4.2484.82% ±2.9990.15% ±0.4993.73% ±0.4894.02% ±0.8560.99% ±2.7886.52% ±0.3992.39% ±0.4183.76% ±1.7182.35% ±1.461.87%2.93%3.39%okp85_not_met –
36 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m2 jhu-clsp/mmBERT-base0.082 m2encoderzephyrus 3 3 111M 17.8 ms 93.32% ±1.0483.47% ±1.3189.80% ±1.3885.57% ±1.6788.54% ±1.8393.79% ±0.8093.60% ±3.6756.84% ±8.2787.78% ±0.3591.54% ±0.8683.20% ±4.8080.14% ±7.6992.17% ±1.2679.20% ±1.5586.37% ±2.2281.45% ±2.0988.22% ±1.9192.75% ±0.9893.73% ±3.5656.07% ±8.3484.49% ±0.6791.49% ±0.9284.05% ±5.1580.07% ±7.821.33%3.47%4.23%okp85_not_met –
37 e2_qwen3_17b_base_cross_ce_real_lora_m2 Qwen/Qwen3-1.7B-Base0.101 m2decoderlegion 3 3 17M 48.0 ms 93.29% ±0.7283.90% ±1.6391.05% ±2.0086.42% ±2.0089.41% ±1.1693.82% ±0.5593.07% ±2.4454.65% ±5.0287.28% ±0.6991.55% ±0.6579.73% ±0.4674.96% ±1.7692.24% ±0.8379.69% ±2.3487.68% ±3.6282.38% ±3.0289.17% ±1.1792.87% ±0.6393.45% ±2.6154.15% ±5.0083.97% ±1.2291.44% ±0.7180.06% ±0.4974.77% ±2.122.40%2.67%4.49%cần sửap85_not_met –
38 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m2 vinai/phobert-base-v20.068 m2encoderzephyrus 6 6 86M 19.6 ms 94.00% ±0.7687.78% ±1.1691.06% ±1.9688.46% ±1.3490.23% ±0.9894.38% ±0.6595.87% ±1.1858.47% ±5.0391.07% ±0.6293.16% ±0.5683.60% ±3.9185.19% ±2.8993.10% ±0.8884.83% ±2.6087.09% ±3.5584.46% ±2.5590.10% ±0.9993.54% ±0.7696.30% ±1.2958.21% ±5.1688.50% ±0.8093.09% ±0.5383.48% ±3.8985.17% ±2.971.07%2.00%3.95%cần sửap85_not_met –
39 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m2 vinai/phobert-base-v20.052 m2encoderzephyrus 3 3 86M 20.9 ms 93.61% ±0.3386.48% ±1.2090.39% ±2.3687.35% ±1.4289.67% ±0.6394.08% ±0.2195.47% ±1.2255.84% ±4.3589.46% ±1.5791.70% ±0.2783.73% ±4.8273.85% ±4.3192.59% ±0.4882.28% ±2.3086.01% ±4.8082.70% ±3.1689.41% ±0.7193.14% ±0.3495.44% ±1.7855.31% ±4.7686.76% ±2.2391.55% ±0.2083.48% ±5.1573.70% ±3.941.33%1.33%4.36%okp85_not_met –
40 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m2 vinai/phobert-base-v20.058 m2encoderzephyrus 3 3 135M 19.6 ms 93.48% ±0.3685.33% ±1.4590.84% ±2.3987.13% ±1.7589.30% ±0.7493.91% ±0.3194.40% ±0.8056.84% ±2.8088.58% ±1.5891.39% ±0.7386.67% ±0.4671.16% ±1.8992.47% ±0.4781.32% ±2.5987.53% ±4.6083.23% ±3.3289.08% ±0.8592.98% ±0.4094.87% ±0.0056.28% ±3.2985.57% ±1.8891.21% ±0.7386.89% ±0.9970.49% ±1.971.60%2.13%4.12%okp85_not_met –
41 e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m2 unsloth/gemma-3-1b-pt0.056 m2decoderlegion 3 3 13M 49.4 ms 93.17% ±0.6284.97% ±1.4990.08% ±1.4586.31% ±0.1488.63% ±1.1593.74% ±0.5294.40% ±0.8053.03% ±3.7187.11% ±0.8791.73% ±0.3580.53% ±3.3378.99% ±3.7992.12% ±0.6281.15% ±1.6986.77% ±2.0082.49% ±0.1688.41% ±1.0292.79% ±0.5295.16% ±1.7852.42% ±3.0983.34% ±0.9591.67% ±0.3780.91% ±3.5678.80% ±3.721.87%2.67%4.82%okp85_not_met –
42 e2_gemma3_270m_cross_ce_real_lora_m2 unsloth/gemma-3-270m0.098 m2decoderzephyrus 3 3 4M 55.8 ms 91.99% ±0.7481.61% ±0.2689.06% ±2.5183.76% ±1.3686.89% ±1.4092.61% ±0.6693.07% ±1.2251.17% ±2.9685.96% ±0.4490.04% ±0.3580.80% ±1.6076.03% ±4.1490.67% ±1.0077.57% ±1.0186.09% ±4.0379.75% ±2.1886.56% ±1.5991.42% ±0.9093.73% ±0.9950.72% ±3.3782.27% ±0.1789.85% ±0.4880.91% ±1.7875.84% ±4.502.67%2.67%5.11%okp85_not_met –
43 e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m2 intfloat/multilingual-e5-large-instruct0.029 m2encoderlegion 3 3 304M 12.6 ms 94.08% ±0.3386.58% ±1.6189.48% ±2.0287.09% ±0.7889.85% ±0.6594.48% ±0.3294.67% ±1.6757.94% ±1.3088.15% ±0.2292.19% ±0.6884.53% ±0.4678.40% ±3.7193.22% ±0.3482.68% ±1.7184.85% ±3.5582.40% ±1.6089.69% ±0.6093.69% ±0.3595.44% ±1.3157.30% ±1.4384.58% ±0.4692.25% ±0.6285.75% ±1.3178.32% ±3.621.07%3.73%3.94%cần sửap85_not_met –
44 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m2 jhu-clsp/mmBERT-base0.040 m2encoderzephyrus 3 3 111M 16.6 ms 92.58% ±1.2483.13% ±2.2788.36% ±1.4584.39% ±1.7587.69% ±2.2893.27% ±1.0495.47% ±0.9250.75% ±4.6188.54% ±1.9892.38% ±0.7884.27% ±2.0182.19% ±2.8691.32% ±1.3578.21% ±1.5683.47% ±1.3079.33% ±1.0187.35% ±2.1992.15% ±1.1395.44% ±0.4949.87% ±4.2585.27% ±2.5592.29% ±0.8384.33% ±1.7881.89% ±3.100.00%2.93%5.35%okp85_not_met –
45 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m2 jhu-clsp/mmBERT-base0.057 m2encoderzephyrus 3 3 111M 19.6 ms 93.29% ±0.9083.87% ±1.7591.00% ±0.4486.11% ±1.2589.10% ±1.1593.85% ±0.7195.73% ±1.2254.48% ±4.9987.72% ±0.4791.78% ±0.3579.20% ±0.8081.59% ±0.4192.33% ±0.6780.41% ±1.7988.09% ±0.4782.67% ±1.4188.92% ±0.9092.97% ±0.5596.01% ±1.3154.35% ±3.5984.89% ±0.9691.64% ±0.4079.20% ±0.4981.29% ±0.370.53%2.13%4.64%okp85_not_met –
46 e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m2 jhu-clsp/mmBERT-base0.070 m2encoderzephyrus 1 1 111M 21.3 ms 94.03%86.08%92.29%88.31%89.68%94.36%94.40%59.90%89.36%91.52%85.60%79.26%93.19%83.00%90.85%85.75%89.57%93.58%94.87%60.00%86.51%91.31%85.47%78.74%0.80%3.20%3.62%okp85_not_met –
47 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m2 jhu-clsp/mmBERT-base0.047 m2encoderzephyrus 3 3 111M 20.2 ms 93.13% ±0.3984.20% ±2.3089.16% ±0.9185.71% ±1.3788.56% ±0.7393.64% ±0.2993.07% ±1.2254.56% ±2.6487.59% ±1.3991.29% ±0.5480.27% ±3.3380.05% ±0.5092.05% ±0.5079.86% ±3.2484.86% ±2.2781.28% ±2.5388.31% ±0.7492.65% ±0.3893.45% ±1.3154.08% ±2.8883.82% ±2.2891.10% ±0.5980.06% ±3.4579.60% ±0.621.07%4.80%4.45%cần sửap85_not_met –
48 e2_xlmr_base_khonly_ce_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.012 m2encoderzephyrus 1 1 86M 7.9 ms 92.38%83.56%88.56%84.89%87.96%92.95%96.00%52.40%87.38%91.11%82.40%81.75%91.13%77.82%82.87%79.08%87.62%91.80%96.58%51.60%84.17%91.00%82.91%81.51%0.80%0.80%4.99%okp85_not_met –
49 e3_ensemble_seed_kd_a05_t1_m2 –0.078 m2ensemblelegion 1 1 N/A 21.2 ms 94.42%89.45%91.47%89.62%91.18%94.74%96.80%60.50%92.43%94.00%85.60%86.99%93.66%86.84%87.36%85.93%91.11%94.03%97.44%60.32%90.18%93.93%85.47%86.96%0.00%2.40%3.62%cần sửap85_not_met –
50 e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m2 unsloth/gemma-3-1b-pt0.042 m2encoderlegion 3 3 100M 55.9 ms 92.76% ±0.3386.59% ±0.7189.73% ±1.4986.92% ±0.5987.99% ±0.5493.38% ±0.2694.93% ±1.6751.46% ±2.4088.99% ±0.7592.01% ±0.7078.67% ±3.9584.76% ±4.2491.71% ±0.3283.52% ±0.7386.38% ±3.0683.51% ±1.4287.81% ±0.5392.41% ±0.2995.44% ±1.7851.37% ±2.0385.58% ±0.6191.85% ±0.6878.35% ±4.0484.63% ±4.511.87%2.93%5.14%okp85_not_met –
51 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m2 vinai/phobert-base-v20.024 m2encoderzephyrus 3 3 86M 19.5 ms 93.03% ±0.3887.20% ±1.1589.63% ±2.6387.11% ±1.8388.80% ±0.8793.63% ±0.3796.00% ±0.8051.89% ±0.9389.43% ±0.6592.13% ±0.5686.67% ±0.9274.89% ±0.9891.94% ±0.4083.49% ±2.6884.84% ±4.1582.52% ±3.5588.54% ±0.8892.63% ±0.4096.01% ±0.4951.31% ±0.8386.47% ±0.7891.97% ±0.5986.61% ±0.9974.51% ±0.851.33%1.87%5.10%okp85_not_met –
52 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m2 vinai/phobert-base-v20.031 m2encoderzephyrus 3 3 86M 22.1 ms 93.30% ±1.0486.15% ±0.8789.34% ±2.9086.63% ±1.8388.83% ±1.7193.80% ±0.8693.60% ±2.1254.30% ±5.6488.68% ±1.0491.41% ±1.2982.67% ±0.9274.55% ±4.5292.24% ±1.1881.59% ±1.0985.08% ±5.1981.94% ±3.1888.55% ±1.7092.83% ±0.9794.02% ±2.2653.41% ±5.4985.75% ±1.1791.17% ±1.3582.34% ±0.9973.77% ±4.730.80%4.27%4.59%okp85_not_met –
53 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t2_tsa_m2 vinai/phobert-base-v20.070 m2encoderzephyrus 3 3 86M 26.8 ms 93.72% ±0.4587.32% ±1.1492.43% ±0.5089.02% ±0.4789.73% ±0.2094.12% ±0.3693.60% ±2.4057.43% ±3.3590.24% ±1.1992.73% ±0.6782.13% ±1.6783.47% ±1.9692.73% ±0.5884.79% ±1.6590.54% ±1.1386.62% ±0.5989.52% ±0.3393.21% ±0.4794.59% ±2.1557.06% ±3.7487.69% ±1.4692.91% ±0.7084.33% ±1.7884.09% ±1.371.60%3.20%4.00%okp85_not_met –
54 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m2 vinai/phobert-base-v20.031 m2encoderzephyrus 3 3 135M 22.5 ms 93.61% ±0.8187.38% ±1.6289.90% ±1.9287.64% ±1.7389.31% ±0.9494.06% ±0.6596.00% ±0.8056.35% ±4.9388.36% ±0.5691.48% ±0.7285.33% ±2.5773.96% ±4.1192.61% ±1.0283.74% ±3.0085.65% ±3.4283.49% ±3.1689.08% ±1.0593.14% ±0.8496.30% ±1.3155.69% ±5.1185.21% ±0.7291.24% ±0.7985.19% ±2.7573.17% ±4.281.07%1.87%4.30%okp85_not_met –
55 e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m2 unsloth/gemma-3-1b-pt0.096 m2decoderlegion 3 3 13M 49.3 ms 93.16% ±0.9481.60% ±2.7486.05% ±4.7882.91% ±3.7889.02% ±1.0293.64% ±0.7192.00% ±0.0055.37% ±6.6286.40% ±0.7391.74% ±0.3583.20% ±2.1276.54% ±2.2992.22% ±1.2277.87% ±2.2683.18% ±1.9879.55% ±2.3888.93% ±1.2092.79% ±0.9392.31% ±0.0055.75% ±7.3082.19% ±0.7891.70% ±0.3483.76% ±2.2676.62% ±2.054.80%2.13%4.33%okp85_not_met –
56 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m2 intfloat/multilingual-e5-large-instruct0.038 m2encoderlegion 3 3 304M 12.9 ms 93.06% ±0.7184.93% ±0.6590.74% ±1.0686.68% ±0.5288.76% ±0.4893.65% ±0.5696.00% ±0.8052.92% ±3.9888.51% ±0.3392.70% ±0.2585.60% ±2.7781.92% ±2.2991.94% ±0.8280.97% ±0.6487.48% ±1.9382.90% ±1.0488.53% ±0.5492.65% ±0.6697.15% ±0.4952.25% ±3.7784.75% ±0.6292.89% ±0.3888.03% ±2.2681.99% ±2.411.07%1.87%4.93%okp85_not_met –
57 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.017 m2encoderzephyrus 3 3 86M 14.2 ms 93.81% ±0.2288.33% ±2.1887.25% ±0.3886.45% ±0.8589.71% ±0.4694.21% ±0.1793.87% ±1.8556.83% ±1.8087.23% ±1.6391.81% ±0.6884.80% ±2.1274.49% ±1.9392.87% ±0.2883.62% ±6.1180.33% ±0.4778.85% ±0.8489.57% ±0.5393.29% ±0.2194.87% ±2.2656.50% ±2.0382.45% ±2.2191.75% ±0.7385.75% ±1.9773.97% ±2.041.60%3.20%4.09%okp85_not_met –
58 e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m2 vinai/phobert-base-v20.026 m2encoderzephyrus 3 3 135M 9.2 ms 94.23% ±0.7587.49% ±0.7488.53% ±1.6987.28% ±1.5590.21% ±1.4294.52% ±0.6292.00% ±1.3961.41% ±6.2188.32% ±0.3291.88% ±0.4783.73% ±4.4176.30% ±1.9793.40% ±0.7982.24% ±2.7383.07% ±3.0881.58% ±3.1490.11% ±1.3793.72% ±0.6793.16% ±1.4860.97% ±5.8884.61% ±0.8191.83% ±0.5184.62% ±5.1375.86% ±2.002.40%3.73%3.37%okp85_not_met –
59 e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.065 m2encoderzephyrus 3 3 87M 20.9 ms 93.39% ±0.3785.45% ±0.8790.23% ±1.3286.90% ±0.7489.13% ±0.6893.86% ±0.2994.40% ±1.6055.44% ±2.8588.01% ±0.4691.13% ±0.4085.33% ±2.4471.03% ±2.0592.38% ±0.5081.40% ±1.1886.41% ±2.1782.87% ±1.2388.92% ±0.7792.93% ±0.3894.59% ±1.3155.05% ±3.4584.48% ±0.4290.90% ±0.3785.19% ±2.6170.26% ±1.921.87%1.60%4.36%okp85_not_met –
60 e2_xlmr_base_cross_logitadjust_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.018 m2encoderzephyrus 3 3 86M 10.6 ms 92.06% ±0.6182.55% ±1.0689.51% ±0.4984.70% ±0.9587.42% ±1.0292.77% ±0.5095.20% ±1.6049.47% ±2.5887.27% ±0.2990.50% ±0.5584.80% ±2.8871.92% ±2.8190.73% ±0.8177.94% ±1.1285.10% ±0.7380.07% ±1.1187.08% ±1.1891.57% ±0.6695.73% ±2.2648.88% ±3.0184.12% ±0.4290.35% ±0.6285.75% ±2.7571.46% ±2.720.80%3.20%5.58%okp85_not_met –
61 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m2 intfloat/multilingual-e5-large-instruct0.026 m2encoderlegion 3 3 304M 13.1 ms 94.53% ±0.6785.85% ±0.6791.92% ±0.6888.12% ±0.8790.93% ±1.0594.86% ±0.5294.13% ±1.6762.23% ±6.6288.56% ±0.0792.78% ±0.1983.47% ±1.2281.55% ±1.6093.61% ±0.7981.89% ±0.6489.00% ±0.6884.44% ±0.8690.66% ±1.0394.02% ±0.6094.02% ±2.2661.31% ±6.5284.84% ±0.0992.75% ±0.2884.05% ±1.3181.31% ±1.681.07%3.47%3.34%okp85_not_met –
62 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m2 jhu-clsp/mmBERT-base0.072 m2encoderzephyrus 3 3 111M 22.0 ms 93.58% ±0.4184.09% ±2.1688.53% ±1.5885.55% ±1.3388.79% ±0.3893.94% ±0.2890.93% ±2.5758.33% ±3.9387.32% ±1.5890.96% ±0.9780.27% ±3.2380.49% ±0.4492.50% ±0.4579.41% ±2.8684.23% ±2.5780.96% ±2.4488.48% ±0.4092.96% ±0.3091.17% ±2.7557.47% ±3.6983.55% ±2.3990.78% ±1.0280.06% ±3.2480.07% ±0.432.13%5.60%3.75%okp85_not_met –
63 e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m2 jhu-clsp/mmBERT-base0.082 m2encoderzephyrus 3 3 111M 20.5 ms 93.38% ±0.1183.52% ±1.3689.37% ±1.7485.38% ±1.3888.80% ±0.5393.82% ±0.1894.67% ±1.6756.47% ±1.2787.26% ±1.6091.11% ±0.7583.47% ±3.3376.67% ±2.7192.27% ±0.2479.07% ±2.0985.54% ±3.1880.99% ±2.3488.52% ±0.6292.82% ±0.2795.16% ±1.7855.89% ±1.4183.57% ±2.5290.93% ±0.7883.48% ±3.4576.26% ±2.971.07%2.40%4.18%okp85_not_met –
64 e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m2 Qwen/Qwen3-1.7B-Base0.067 m2encoderlegion 1 1 105M 48.8 ms 91.99%84.66%90.53%86.20%87.44%92.82%93.60%47.18%88.43%91.83%83.20%78.20%90.71%81.16%87.53%82.86%87.12%91.68%94.02%46.61%84.75%91.70%82.91%78.23%2.40%3.20%6.00%okp85_not_met –
65 e2_visobert_cross_ce_real_frozenemb_lr4e5_m2 uitnlp/visobert0.037 m2encoderzephyrus 1 1 86M 18.8 ms 89.96%81.79%87.60%82.82%84.98%91.24%94.40%40.83%86.16%88.58%88.00%59.78%88.65%78.51%83.19%78.96%84.76%90.07%94.87%40.96%83.14%88.34%88.89%59.09%2.40%2.40%7.83%okp85_not_met –
66 e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m2 FacebookAI/xlm-roberta-base0.027 m2encoderzephyrus 1 1 86M 7.8 ms 92.03%83.16%89.17%85.11%87.73%92.71%90.40%49.56%86.74%90.30%82.40%69.59%90.87%78.62%84.93%80.64%87.54%91.64%91.45%49.54%82.91%90.08%82.91%68.79%3.20%4.80%5.26%okp85_not_met –
67 e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m2 jhu-clsp/mmBERT-base0.057 m2encoderzephyrus 3 3 110M 36.2 ms 93.13% ±0.9188.90% ±1.6388.12% ±1.3886.97% ±1.0889.15% ±1.2393.70% ±0.7096.00% ±1.6053.35% ±5.2088.72% ±1.8491.92% ±0.9879.20% ±3.6779.09% ±5.4492.20% ±0.8486.34% ±3.7282.51% ±2.6381.27% ±2.2289.02% ±1.0692.81% ±0.6596.58% ±1.7153.24% ±4.1684.88% ±2.5491.75% ±1.1278.92% ±3.8578.58% ±5.851.07%2.67%4.88%okp85_not_met –
68 e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.034 m2encoderzephyrus 3 3 87M 20.6 ms 94.08% ±0.3087.24% ±2.3690.37% ±2.1687.96% ±0.6990.15% ±0.5594.39% ±0.2192.53% ±1.2260.55% ±4.0090.07% ±2.2192.15% ±0.5085.60% ±1.3977.84% ±3.1393.08% ±0.4884.16% ±4.8086.01% ±4.5083.35% ±2.1789.82% ±0.7593.44% ±0.3692.88% ±1.3159.96% ±4.3987.43% ±3.0491.86% ±0.5185.47% ±1.4877.45% ±3.321.60%4.27%3.48%okp85_not_met –
69 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m2 jhu-clsp/mmBERT-base0.144 m2encoderzephyrus 3 3 111M 18.3 ms 93.04% ±0.5683.57% ±0.1989.00% ±1.4885.26% ±0.9288.49% ±0.7193.57% ±0.4293.60% ±0.8054.35% ±3.5887.87% ±1.7891.18% ±1.2884.27% ±4.6274.71% ±9.5391.83% ±0.6578.71% ±0.9884.64% ±2.8080.45% ±1.8888.13% ±0.7792.49% ±0.5093.73% ±1.3153.16% ±3.5183.80% ±3.0291.01% ±1.4284.33% ±4.7174.06% ±9.950.53%4.00%4.53%okp85_not_met –
70 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m2 jhu-clsp/mmBERT-base0.108 m2encoderzephyrus 3 3 111M 18.7 ms 94.08% ±0.4685.38% ±2.1090.19% ±0.4187.03% ±1.3689.72% ±0.5394.40% ±0.3093.60% ±2.1261.36% ±5.7588.82% ±1.2392.05% ±0.5586.13% ±3.6182.46% ±1.6793.12% ±0.5181.28% ±2.5086.87% ±1.2983.14% ±1.9489.47% ±0.4993.51% ±0.3294.02% ±2.2660.52% ±5.6585.46% ±1.2091.94% ±0.5386.32% ±3.4282.19% ±1.811.33%4.53%3.43%okp85_not_met –
71 e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m2 Qwen/Qwen3-0.6B-Base0.075 m2decoderlegion 3 3 10M 45.9 ms 93.06% ±1.2383.85% ±1.2792.24% ±1.6986.68% ±1.6689.26% ±1.9293.67% ±0.9496.00% ±1.3953.36% ±6.7388.57% ±0.9492.82% ±0.6884.00% ±3.4981.53% ±2.9391.94% ±1.5380.33% ±1.6389.71% ±2.5283.32% ±2.0888.96% ±2.0592.66% ±1.1896.01% ±1.3152.78% ±6.9685.67% ±1.2192.79% ±0.6684.33% ±3.2481.68% ±2.880.53%2.40%4.91%okp85_not_met –
72 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t1_tsa_m2 vinai/phobert-base-v20.070 m2encoderzephyrus 3 3 86M 22.7 ms 93.51% ±0.4186.98% ±1.5391.91% ±1.9188.47% ±1.2889.40% ±0.3993.95% ±0.3395.47% ±1.6756.35% ±3.2490.16% ±1.2192.93% ±0.5584.00% ±1.6085.22% ±2.6092.43% ±0.5284.07% ±2.4989.47% ±3.6285.64% ±2.5989.14% ±0.4892.95% ±0.4295.73% ±1.4855.81% ±3.5387.29% ±1.6293.04% ±0.5685.75% ±1.3185.36% ±2.781.33%1.60%4.26%okp85_not_met –
73 e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m2 Qwen/Qwen3-0.6B-Base0.077 m2encoderlegion 3 3 97M 50.7 ms 91.54% ±1.1785.72% ±2.2689.29% ±1.8486.20% ±2.4687.68% ±1.2092.35% ±0.8792.27% ±3.2347.85% ±5.3188.92% ±1.6491.34% ±0.0377.87% ±2.0177.86% ±1.7190.73% ±1.9182.40% ±4.4985.49% ±3.6482.69% ±4.5887.85% ±1.7791.53% ±1.4692.59% ±3.4550.09% ±8.0185.94% ±2.4491.16% ±0.0677.78% ±2.5677.77% ±1.902.93%4.27%5.87%cần sửap85_not_met –
74 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m2 vinai/phobert-base-v20.104 m2encoderzephyrus 3 3 86M 22.0 ms 93.98% ±1.0987.91% ±1.0390.08% ±2.3388.10% ±1.5889.93% ±2.0794.31% ±0.8393.33% ±3.2361.64% ±9.1390.16% ±0.5592.36% ±0.3286.93% ±1.6780.25% ±4.1093.00% ±1.2585.38% ±2.9086.17% ±3.9484.08% ±2.6689.65% ±2.0993.37% ±0.9792.88% ±3.4561.18% ±9.2687.24% ±0.5592.22% ±0.3286.61% ±1.3180.34% ±3.762.93%2.40%3.49%okp85_not_met –
75 e2_xlmr_base_cross_ce_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.044 m2encoderzephyrus 3 3 86M 12.1 ms 92.45% ±1.2983.47% ±0.9588.23% ±2.2184.79% ±2.0587.92% ±1.7293.03% ±1.0191.73% ±3.2353.13% ±8.8386.40% ±2.2190.24% ±1.2382.40% ±6.0472.01% ±11.1491.25% ±1.5877.72% ±2.8283.06% ±4.3979.16% ±3.9987.68% ±1.9191.93% ±1.2592.59% ±3.2452.71% ±9.3782.29% ±3.6390.13% ±1.4183.19% ±5.5671.73% ±11.642.40%4.80%4.84%okp85_not_met –
76 e2_xlmr_base_cross_wcesqrt_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.019 m2encoderzephyrus 3 3 86M 10.9 ms 92.61% ±0.8383.49% ±1.3290.95% ±1.2386.07% ±1.3588.29% ±1.1093.15% ±0.5794.40% ±2.8854.14% ±6.0986.77% ±0.1491.09% ±0.4882.93% ±4.0376.62% ±7.5491.32% ±1.0179.48% ±1.5688.02% ±2.6382.31% ±1.9987.92% ±1.2191.98% ±0.6995.16% ±2.1553.26% ±6.1782.94% ±0.7791.02% ±0.6184.05% ±3.5676.41% ±7.721.07%3.47%4.68%okp85_not_met –
77 e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.049 m2encoderzephyrus 3 3 86M 18.9 ms 93.28% ±0.2584.06% ±1.0190.27% ±1.6086.10% ±0.9988.98% ±0.4493.77% ±0.3095.20% ±0.8054.85% ±0.9587.93% ±0.6091.60% ±0.4085.87% ±3.0374.07% ±1.4492.26% ±0.2579.59% ±1.4786.46% ±3.2081.79% ±2.0288.78% ±0.3992.84% ±0.3095.44% ±0.9954.50% ±1.3784.72% ±0.7191.45% ±0.3386.04% ±3.5673.72% ±1.681.60%1.07%4.49%okp85_not_met –
78 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m2 vinai/phobert-base-v20.159 m2encoderzephyrus 3 3 86M 22.1 ms 93.48% ±1.0385.33% ±2.9387.11% ±5.0585.07% ±5.0889.40% ±1.2793.81% ±0.8192.80% ±2.1259.12% ±8.4589.01% ±1.1491.11% ±0.5086.93% ±3.7872.22% ±2.0292.59% ±1.0982.16% ±1.7685.22% ±2.4082.81% ±2.5589.27% ±1.1992.99% ±0.8293.45% ±2.4758.97% ±7.7885.96% ±1.8291.02% ±0.4487.75% ±4.0471.89% ±1.922.40%3.73%3.81%okp85_not_met –
79 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsa_m2 vinai/phobert-base-v20.085 m2encoderzephyrus 3 3 86M 22.9 ms 93.92% ±0.8085.85% ±1.4191.27% ±2.4387.73% ±1.9690.07% ±1.2694.28% ±0.6594.13% ±2.4459.44% ±5.3490.22% ±1.2992.70% ±0.7483.20% ±2.8883.63% ±1.4692.94% ±0.9581.66% ±2.3687.93% ±4.6183.78% ±3.4789.85% ±1.3393.36% ±0.7894.30% ±2.1558.99% ±5.3987.38% ±1.5892.54% ±0.7982.91% ±3.0883.37% ±1.561.07%2.40%3.74%okp85_not_met –
80 e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m2 Qwen/Qwen3-Embedding-4B0.061 m2decoderkaggle 2 2 79M 126.5 ms 91.08% ±0.2482.85% ±4.1887.24% ±2.4783.16% ±1.6085.49% ±1.0191.68% ±0.3492.80% ±3.3950.76% ±0.7686.19% ±1.7488.11% ±2.1872.00% ±13.5879.62% ±7.3089.39% ±0.3079.60% ±4.5383.92% ±4.5079.14% ±1.2984.81% ±1.0590.14% ±0.4992.31% ±3.6349.53% ±1.3083.66% ±2.6087.62% ±2.3771.37% ±13.9079.17% ±7.822.00%3.60%5.15%okp85_not_met –
81 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m2 vinai/phobert-base-v20.025 m2encoderzephyrus 1 1 86M 11.5 ms 93.59%81.37%78.94%78.16%88.79%93.79%90.40%58.85%83.65%90.01%79.20%75.00%92.87%78.67%79.02%77.79%88.90%93.14%91.45%60.11%77.19%89.96%80.34%74.60%4.00%4.80%3.62%okp85_not_met –
82 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m2 Qwen/Qwen3-Embedding-0.6B0.152 m2decoderzephyrus 3 3 13M 62.5 ms 91.79% ±0.8281.06% ±0.7388.11% ±0.5683.46% ±0.5586.20% ±1.4792.20% ±0.7186.67% ±0.4655.34% ±5.4685.17% ±1.0087.82% ±0.4676.00% ±2.7771.59% ±0.9590.25% ±1.0377.02% ±0.4985.36% ±0.9179.54% ±0.2385.61% ±1.6190.79% ±0.9086.04% ±0.4954.02% ±5.8081.79% ±1.7087.35% ±0.4875.21% ±2.9670.76% ±1.296.40%3.73%4.06%okp85_not_met –
83 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m2 vinai/phobert-base-v20.035 m2encoderzephyrus 3 3 86M 20.7 ms 93.61% ±1.0086.06% ±0.0790.72% ±0.6887.43% ±0.6889.57% ±1.0794.09% ±0.7795.47% ±1.2256.20% ±6.1189.69% ±0.3492.32% ±0.8784.00% ±4.8078.31% ±3.8392.56% ±1.1582.01% ±0.3586.87% ±1.5483.36% ±1.0489.26% ±1.1093.14% ±0.8895.73% ±1.4855.26% ±6.0386.87% ±0.7092.13% ±0.9383.76% ±5.1377.74% ±4.041.07%1.60%4.35%cần sửap85_not_met –
84 e2_me5_base_cross_ce_real_frozenemb_lr1e5_m2 intfloat/multilingual-e5-base0.052 m2encoderzephyrus 1 1 86M 11.1 ms 92.29%80.10%80.62%78.62%87.68%92.83%92.00%51.11%83.56%88.84%79.20%63.46%91.29%78.38%81.82%79.02%87.55%91.93%92.31%51.67%78.20%88.84%81.20%63.33%2.40%4.80%5.03%okp85_not_met –
85 e2_visobert_cross_ce_real_frozenemb_lr2e5_m2 uitnlp/visobert0.027 m2encoderzephyrus 1 1 86M 17.1 ms 90.17%80.29%84.65%80.98%85.52%91.15%91.20%43.51%84.50%89.09%79.20%68.28%88.75%73.21%77.54%73.81%85.24%89.86%92.31%43.37%78.84%88.89%79.49%67.88%4.00%4.00%6.77%okp85_not_met –
86 e2_visobert_cross_ce_real_frozenemb_lr1e5_m2 uitnlp/visobert0.186 m2encoderzephyrus 1 1 86M 12.5 ms 90.13%79.34%84.01%80.42%84.20%90.93%87.20%45.61%81.43%85.40%83.20%51.74%88.60%73.89%77.95%74.59%83.79%89.51%87.18%45.33%75.74%85.17%83.76%51.85%5.60%4.80%5.95%okp85_not_met –
87 e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m2 FacebookAI/xlm-roberta-base0.161 m2encoderzephyrus 1 1 86M 12.5 ms 91.95%79.30%85.31%81.36%87.26%92.41%91.20%54.55%82.07%88.11%88.00%60.11%90.60%73.09%78.29%74.69%86.97%91.18%92.31%54.00%75.47%87.90%88.89%59.77%3.20%4.80%4.35%okp85_not_met –
88 e1_tfidf_hybrid_qwen3emb_lr_m2 Qwen/Qwen3-Embedding-0.6B0.149 m2tfidf_hybridzephyrus 1 1 157,386 35.1 ms 90.52%82.54%83.09%81.85%84.30%91.05%85.60%49.31%84.97%87.15%74.40%67.88%89.02%76.74%77.20%75.98%83.91%89.59%85.47%49.26%81.05%86.59%73.50%66.67%6.40%5.60%5.03%okp85_not_met –
89 e1_tfidf_svmcal_ce_real_base_m2 TF-IDF0.178 m2tfidfzephyrus 1 1 111,636 4.5 ms 92.25%84.45%85.09%84.24%87.18%92.55%84.80%57.92%86.94%88.95%72.80%79.13%90.92%78.63%79.26%78.39%86.79%91.26%84.62%57.56%83.27%88.49%71.79%78.50%9.60%2.40%3.52%okp85_not_met –
90 e2_dual_qwen3emb06b_concat_ce_real_frozen_m2 Qwen/Qwen3-Embedding-0.6B0.227 m2decoderzephyrus 3 3 2M 54.5 ms 89.06% ±0.8184.35% ±1.9981.52% ±2.4881.27% ±1.0782.08% ±1.0889.63% ±0.6787.20% ±0.8047.17% ±3.5183.48% ±1.3884.30% ±1.1181.07% ±0.4661.28% ±3.2287.19% ±0.9081.78% ±3.7475.46% ±3.6775.26% ±2.4681.44% ±1.0887.82% ±0.7386.89% ±0.4946.39% ±3.4178.45% ±2.3483.84% ±1.2381.48% ±0.4960.71% ±3.465.87%5.60%5.63%okp85_not_met –
– e3_fbverifier_devtop1_familygeomean_fbf1_m2 –hard m2ensemblezephyrus 1 1 N/A 271.8 ms 96.75%89.47%93.85%91.26%94.34%96.77%N/A87.20%91.26%94.34%87.20%87.20%96.41%86.23%92.69%88.52%94.46%96.46%N/A87.29%88.52%94.46%88.03%87.29%4.00%5.60%0.73%ok –
– e3_fbverifier_devtop1_familygeomean_macrof1_m2 –hard m2ensemblezephyrus 1 1 N/A 271.8 ms 96.75%89.47%93.85%91.26%94.34%96.77%N/A87.20%91.26%94.34%87.20%87.20%96.41%86.23%92.69%88.52%94.46%96.46%N/A87.29%88.52%94.46%88.03%87.29%4.00%5.60%0.73%ok –
1 e3_ensemble_devtop1_familygeomean_m1ensemble trung bình nhân theo nhóm trên 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051 m1ensemblezephyrus 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 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051 m1ensemblezephyrus 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 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051 m1ensemblezephyrus 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_m1ensemble trung bình đều theo nhóm trên 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.112 m1ensemblezephyrus 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 2 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.110 m1ensemblezephyrus 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 3 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051 m1ensemblezephyrus 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 4 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.058 m1ensemblezephyrus 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.111 m1ensemblezephyrus 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.087 m1ensemblezephyrus 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 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.087 m1ensemblezephyrus 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.087 m1ensemblezephyrus 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_m1ensemble trung bình nhân trên 5 phương pháp · nhãn m1 –0.073 m1ensemblezephyrus 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_m1ensemble trung bình trên 5 phương pháp · nhãn m1 –0.073 m1ensemblezephyrus 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_m1gemma-3-1b-pt · 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 unsloth/gemma-3-1b-pt0.064 m1decoderlegion 3 3 24M 94.0 ms 92.01% ±1.5484.46% ±1.7387.32% ±0.7084.53% ±1.1287.32% ±2.1492.69% ±1.1394.62% ±0.7751.43% ±7.9387.75% ±0.5090.62% ±0.5980.00% ±4.0080.37% ±2.9990.71% ±1.7478.88% ±0.9681.56% ±1.5278.50% ±1.0586.97% ±2.1791.49% ±1.2795.08% ±0.8250.78% ±7.7583.88% ±1.0290.44% ±0.5479.78% ±4.5180.17% ±2.882.82%2.31%5.40%cần sửap85_not_met –
15 e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-base0.017 m1encoderzephyrus 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_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.155 m1encoderzephyrus 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_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-1.7B-Base0.081 m1decoderlegion 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_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-1.7B-Base0.105 m1decoderlegion 3 3 17M 52.8 ms 92.66% ±0.3185.55% ±0.9592.30% ±0.2987.81% ±0.3687.99% ±0.2993.28% ±0.2993.33% ±0.4451.49% ±0.4387.67% ±1.0090.64% ±0.4683.85% ±1.3372.96% ±5.2691.38% ±0.2882.85% ±1.0690.52% ±0.2685.44% ±0.4987.51% ±0.2892.11% ±0.2692.90% ±0.4750.68% ±0.7884.27% ±1.0290.32% ±0.4983.88% ±1.7172.38% ±5.212.31%3.08%5.13%okp85_not_met –
19 e3_ensemble_top_greedy_m1ensemble chọn tham lam Caruana trên 5 phương pháp · nhãn m1 –0.113 m1ensemblezephyrus 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_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 3e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.017 m1encoderlegion 3 3 304M 12.3 ms 92.86% ±0.8984.13% ±0.3690.35% ±0.7386.18% ±0.3588.36% ±1.5493.33% ±0.6794.36% ±0.4456.58% ±7.5686.40% ±0.8190.59% ±0.1382.82% ±1.1876.43% ±2.6191.66% ±1.0180.29% ±0.6387.35% ±0.8182.53% ±0.2187.99% ±1.5692.23% ±0.7694.54% ±0.9555.64% ±7.4282.56% ±0.8090.43% ±0.0883.33% ±1.7175.96% ±2.551.79%2.56%4.34%okp85_not_met –
21 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m1 jhu-clsp/mmBERT-base0.094 m1encoderzephyrus 3 3 111M 20.7 ms 92.03% ±1.4484.34% ±1.3887.96% ±3.1084.81% ±2.8287.26% ±1.8292.70% ±1.0894.10% ±3.5551.61% ±8.9187.63% ±0.1690.09% ±0.4478.97% ±1.9476.23% ±0.9190.65% ±1.7378.76% ±3.5383.07% ±6.0879.33% ±5.3886.85% ±1.9591.44% ±1.3194.26% ±3.5750.79% ±9.2484.52% ±0.1289.99% ±0.4279.78% ±1.7176.04% ±1.072.05%2.82%5.39%okp85_not_met –
22 e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1factorized · train thật + synth ×1, nhãn m1 jhu-clsp/mmBERT-base0.108 m1encoderzephyrus 3 3 111M 17.0 ms 90.57% ±1.3082.11% ±0.5388.67% ±2.4983.77% ±1.9085.01% ±1.3991.49% ±0.9092.56% ±5.5646.27% ±6.8086.32% ±0.9988.64% ±2.3386.41% ±6.4168.16% ±13.0389.39% ±1.3578.36% ±1.2685.55% ±4.8380.16% ±3.0584.82% ±1.2890.36% ±0.9592.08% ±5.9346.89% ±6.3883.11% ±1.1988.45% ±2.2786.61% ±6.3768.00% ±12.263.08%3.85%6.49%oktest_informedp85_not_met –
23 e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-4B-Base0.056 m1decoderlegion 3 3 33M 138.1 ms 93.00% ±0.3583.41% ±0.3390.97% ±0.7486.01% ±0.1988.90% ±0.3893.51% ±0.2795.38% ±0.7754.20% ±1.9388.04% ±0.1291.56% ±0.5980.00% ±3.3579.79% ±4.6891.85% ±0.3779.17% ±0.3087.35% ±1.4381.84% ±0.5888.46% ±0.4592.46% ±0.2895.08% ±0.8253.35% ±1.9384.49% ±0.3591.41% ±0.6880.05% ±3.1080.06% ±5.051.54%1.79%4.71%okp85_not_met –
24 e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 5 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.045 m1decoderlegion 2 1 lỗi 2 13M 51.1 ms 92.15% ±0.1284.98% ±2.7288.75% ±1.3185.57% ±0.5687.30% ±0.7292.83% ±0.0995.00% ±0.5450.21% ±1.1587.49% ±0.0390.71% ±0.7985.00% ±2.7275.17% ±0.2490.83% ±0.2981.27% ±3.9484.37% ±1.4081.29% ±0.9186.93% ±0.8791.63% ±0.2395.49% ±0.5849.49% ±1.6484.08% ±0.3090.56% ±0.8385.66% ±2.9074.64% ±0.262.69%2.31%5.50%cần sửap85_not_met –
25 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-v20.033 m1encoderzephyrus 3 3 86M 19.2 ms 91.46% ±0.3582.67% ±1.6886.91% ±0.8783.55% ±1.2486.35% ±0.0792.18% ±0.2391.79% ±1.6048.47% ±2.3287.90% ±1.2690.80% ±0.7483.33% ±1.1877.76% ±0.9690.22% ±0.4677.05% ±3.1781.12% ±2.3777.76% ±2.7986.10% ±0.1591.02% ±0.3191.80% ±1.4248.25% ±2.5484.90% ±1.5090.58% ±0.7983.06% ±1.2577.17% ±1.013.85%3.08%5.72%cần sửatest_informedp85_not_met –
26 e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=32 · CE · lr 0.0001, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.045 m1decoderlegion 3 3 26M 49.5 ms 91.42% ±1.0982.69% ±2.0489.35% ±1.9084.63% ±2.1586.62% ±1.6692.19% ±0.7894.10% ±0.8948.73% ±5.8285.59% ±0.7690.05% ±0.2980.00% ±2.7777.81% ±1.3689.98% ±1.3978.82% ±3.0085.85% ±2.8180.72% ±3.0386.23% ±1.8090.92% ±1.0194.54% ±0.9547.68% ±6.0581.76% ±0.6989.88% ±0.0680.05% ±1.7177.34% ±1.872.82%2.56%5.88%okp85_not_met –
27 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 6 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.052 m1encoderzephyrus 3 3 111M 18.0 ms 91.72% ±0.7181.07% ±0.9990.20% ±1.2883.89% ±0.9186.38% ±0.8392.34% ±0.5791.03% ±1.6051.25% ±3.9885.64% ±0.4689.96% ±0.6081.03% ±4.8975.90% ±4.1290.54% ±0.7177.82% ±0.8988.55% ±2.5180.76% ±1.1286.13% ±0.8191.25% ±0.6090.98% ±1.4251.33% ±3.3782.46% ±1.0989.74% ±0.6881.15% ±4.9975.32% ±4.382.56%4.62%5.10%okp85_not_met –
28 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-base0.051 m1encoderzephyrus 3 3 111M 18.6 ms 92.56% ±0.7383.95% ±1.1090.04% ±0.7385.91% ±1.2087.60% ±0.6993.10% ±0.5691.79% ±2.9153.27% ±4.3086.61% ±1.3089.84% ±0.5677.95% ±2.9174.88% ±1.8891.35% ±0.7780.25% ±1.4187.49% ±1.5282.53% ±1.6987.23% ±0.6692.00% ±0.5991.80% ±2.8452.44% ±3.9283.40% ±1.4689.65% ±0.5978.14% ±3.1074.29% ±1.982.31%4.36%4.76%oktest_informedp85_not_met –
29 e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 jhu-clsp/mmBERT-base0.118 m1encoderzephyrus 3 3 111M 18.7 ms 92.17% ±0.9483.31% ±1.9889.49% ±1.5185.21% ±1.9387.05% ±1.0592.77% ±0.6692.82% ±3.7952.13% ±6.7486.87% ±0.1489.83% ±0.7482.56% ±1.7873.00% ±6.3390.83% ±1.1679.43% ±2.6086.45% ±2.8281.54% ±2.8086.64% ±1.1491.56% ±0.8192.90% ±3.8751.28% ±7.0183.60% ±0.4489.55% ±0.8182.51% ±1.7172.25% ±6.612.05%4.36%5.12%okp85_not_met –
30 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1PhoBERT-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 m1 vinai/phobert-base-v20.008 m1encoderzephyrus 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 –
31 e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1 Qwen/Qwen3-4B-Base0.162 m1decoderkaggle 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 –
32 e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1 Qwen/Qwen3-8B-Base0.170 m1decoderkaggle 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 –
33 e3_ensemble_seed6_kd_a05_t1_m1ensemble trung bình nhân trên 1 phương pháp · nhãn m1 –0.067 m1ensemblelegion 1 1 N/A 19.6 ms 93.30%87.51%89.86%87.74%88.92%93.68%94.62%57.48%90.20%91.94%83.85%83.21%92.22%84.77%85.59%83.88%88.66%92.66%95.08%56.86%87.58%91.76%83.61%82.93%1.54%3.08%4.08%okp85_not_met –
34 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_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-v20.073 m1encoderzephyrus 3 3 86M 20.4 ms 92.28% ±0.2682.99% ±0.3890.19% ±1.9585.28% ±0.8087.49% ±0.6992.79% ±0.2194.87% ±0.4453.03% ±1.2786.94% ±1.2890.36% ±0.3883.59% ±1.7874.62% ±2.0091.09% ±0.2278.80% ±0.5386.76% ±4.0881.12% ±2.0587.20% ±0.6091.67% ±0.2095.08% ±0.0052.74% ±1.0683.82% ±1.7890.07% ±0.3783.33% ±1.8974.07% ±2.141.54%2.05%4.91%oktest_informedp85_not_met –
35 e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 unsloth/gemma-3-1b-pt0.064 m1decoderlegion 3 3 13M 49.8 ms 92.55% ±0.0983.40% ±0.7288.72% ±1.4285.09% ±0.9487.87% ±0.5893.03% ±0.1893.33% ±2.9154.46% ±1.4086.55% ±1.4590.61% ±0.7584.36% ±2.7077.42% ±2.0391.31% ±0.1178.76% ±0.8784.34% ±1.4280.48% ±1.0487.52% ±0.5791.89% ±0.2293.44% ±2.8453.74% ±1.3582.28% ±1.3790.50% ±0.7684.97% ±2.8877.17% ±1.833.59%2.82%4.56%okp85_not_met –
36 e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-base0.031 m1encoderzephyrus 3 3 86M 9.0 ms 91.91% ±0.9482.49% ±0.1786.97% ±0.5083.55% ±0.4886.93% ±1.4492.52% ±0.7395.13% ±1.7852.23% ±5.4184.34% ±0.5888.80% ±1.1181.79% ±3.1169.27% ±6.8890.44% ±1.1076.05% ±1.0480.56% ±1.0777.04% ±0.7086.38% ±1.5091.16% ±0.8494.81% ±1.8951.03% ±5.5379.90% ±1.0288.60% ±1.0983.06% ±4.0468.55% ±6.732.82%1.28%5.16%okp85_not_met –
37 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.061 m1encoderzephyrus 3 3 111M 17.6 ms 91.70% ±0.8483.02% ±1.1988.39% ±0.7084.42% ±0.9286.73% ±1.0292.37% ±0.6293.59% ±1.1850.36% ±5.3886.10% ±1.6989.83% ±1.0079.49% ±3.1175.00% ±6.0690.41% ±0.9778.70% ±1.5484.43% ±1.1480.16% ±1.1586.38% ±1.0791.19% ±0.7393.72% ±1.2549.81% ±5.0082.59% ±2.1089.57% ±1.1279.51% ±2.9674.23% ±6.652.31%2.82%5.48%okp85_not_met –
38 e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1 Qwen/Qwen3-4B-Base0.131 m1decoderkaggle 3 3 33M 232.8 ms 93.62% ±0.8584.89% ±2.3291.85% ±0.9487.26% ±0.7789.57% ±1.0993.99% ±0.6494.36% ±2.7060.20% ±9.1688.39% ±0.7391.66% ±0.5682.56% ±2.9180.69% ±0.7792.58% ±0.9881.71% ±3.2689.31% ±1.7883.97% ±1.3989.25% ±1.1093.04% ±0.7693.99% ±2.8859.47% ±9.2085.10% ±0.7591.54% ±0.5382.51% ±2.6380.52% ±0.882.31%1.54%3.80%okp85_not_met –
39 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_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 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.043 m1encoderzephyrus 3 3 86M 20.9 ms 91.93% ±0.4684.32% ±3.0688.06% ±1.9684.91% ±1.2387.18% ±0.7392.51% ±0.3692.05% ±2.4751.47% ±2.3986.78% ±0.4489.86% ±0.1884.36% ±1.1870.91% ±1.0490.61% ±0.4880.94% ±5.9183.16% ±3.5979.93% ±2.1886.81% ±0.6891.28% ±0.3891.80% ±2.1750.77% ±2.3383.06% ±0.9589.56% ±0.2083.88% ±0.9570.42% ±1.004.62%2.31%5.09%okp85_not_met –
40 e2_gemma3_1b_pt_cross_ce_real_lora_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.053 m1decoderlegion 3 3 13M 49.8 ms 92.25% ±0.7383.82% ±1.4888.29% ±0.6284.98% ±0.9587.35% ±0.9992.81% ±0.6493.85% ±0.0052.69% ±2.8786.92% ±1.8490.49% ±0.6282.05% ±1.9479.51% ±2.8591.09% ±0.7879.48% ±1.8084.03% ±0.6080.59% ±0.9887.14% ±0.9291.72% ±0.7094.54% ±0.4752.60% ±2.6483.43% ±2.7590.38% ±0.7082.51% ±2.0679.58% ±2.853.08%2.56%4.94%okp85_not_met –
41 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_kd_a05_t1_tsa_m1 vinai/phobert-base-v20.056 m1encoderzephyrus 3 3 86M 22.1 ms 91.98% ±0.3583.21% ±1.1490.67% ±0.6085.65% ±0.5487.26% ±0.5892.58% ±0.3794.62% ±1.3351.62% ±1.0287.91% ±0.7290.82% ±0.5783.59% ±1.1878.22% ±1.7490.66% ±0.3979.58% ±1.0188.12% ±1.3082.28% ±0.3286.90% ±0.5891.37% ±0.4194.81% ±1.2550.96% ±1.1584.76% ±0.8290.58% ±0.5883.33% ±1.2577.65% ±1.822.31%1.79%5.18%okp85_not_met –
42 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-v20.037 m1encoderzephyrus 3 3 86M 24.1 ms 92.21% ±0.4785.37% ±1.9490.06% ±1.5086.57% ±0.1287.40% ±0.4592.72% ±0.3792.82% ±1.7853.11% ±2.2288.80% ±0.7790.97% ±0.4082.05% ±2.4779.20% ±2.0490.97% ±0.5983.02% ±3.6987.05% ±2.9883.38% ±0.3787.07% ±0.5591.56% ±0.4992.90% ±2.3752.59% ±2.4186.14% ±0.7590.71% ±0.3781.69% ±2.6378.88% ±1.822.56%3.33%4.79%okp85_not_met –
43 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_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, nhãn m1 vinai/phobert-base-v20.085 m1encoderzephyrus 3 3 86M 20.7 ms 92.92% ±0.3984.36% ±1.4789.91% ±2.2286.33% ±0.9388.16% ±0.3593.25% ±0.3691.54% ±0.7758.20% ±2.0887.41% ±1.1490.57% ±0.3182.31% ±1.5479.65% ±1.2091.81% ±0.5780.36% ±2.1886.51% ±4.3282.44% ±2.5587.90% ±0.4992.19% ±0.5391.80% ±0.8257.72% ±2.5484.05% ±1.0790.31% ±0.3681.97% ±1.6479.15% ±1.282.82%4.10%3.85%okp85_not_met –
44 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_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, nhãn m1 vinai/phobert-base-v20.028 m1encoderzephyrus 3 3 86M 22.9 ms 91.82% ±0.3281.49% ±1.7889.13% ±3.4183.85% ±2.3086.93% ±0.0792.46% ±0.2593.08% ±1.3350.22% ±1.2685.81% ±2.1091.10% ±0.7580.77% ±2.6680.88% ±3.8090.49% ±0.3976.90% ±3.6185.66% ±6.8579.38% ±4.6486.56% ±0.1191.27% ±0.3193.17% ±0.9549.43% ±1.0182.26% ±2.8690.91% ±0.7880.33% ±2.8480.87% ±3.762.82%2.56%5.39%okp85_not_met –
45 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.082 m1encoderzephyrus 3 3 111M 17.8 ms 92.00% ±1.0481.32% ±1.2488.43% ±1.3183.60% ±1.4786.67% ±1.6892.55% ±0.7991.79% ±3.8753.54% ±7.3785.66% ±0.5889.51% ±1.0081.54% ±4.6275.03% ±7.3390.59% ±1.2677.00% ±1.3084.89% ±2.1679.32% ±1.8086.22% ±1.7491.27% ±0.9791.80% ±3.7652.68% ±7.3982.24% ±0.9389.32% ±1.0582.24% ±4.9474.70% ±7.422.56%4.10%4.80%okp85_not_met –
46 e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1, nhãn m1 jhu-clsp/mmBERT-base0.057 m1encoderzephyrus 3 3 111M 19.4 ms 91.52% ±0.3581.51% ±1.7289.89% ±0.7784.00% ±1.1086.21% ±0.3392.29% ±0.3192.56% ±2.3548.07% ±1.1085.04% ±1.2489.32% ±0.6880.77% ±3.5370.73% ±5.4090.18% ±0.4977.92% ±1.8887.76% ±1.5680.58% ±1.5585.87% ±0.4491.09% ±0.4192.62% ±2.1747.62% ±1.4781.64% ±1.3489.08% ±0.7381.15% ±3.7670.00% ±5.572.31%3.33%5.83%oktest_informedp85_not_met –
47 e2_qwen3_17b_base_cross_ce_real_lora_m1Qwen3-1.7B Base · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật, nhãn m1 Qwen/Qwen3-1.7B-Base0.101 m1decoderlegion 3 3 17M 48.0 ms 92.01% ±0.6782.36% ±1.6089.72% ±2.0184.86% ±2.0287.54% ±1.1992.62% ±0.4791.54% ±2.7751.55% ±4.4685.64% ±0.5489.59% ±0.5978.21% ±0.4470.44% ±0.5190.71% ±0.7778.09% ±2.3086.25% ±3.6380.72% ±3.0387.18% ±1.2091.44% ±0.5491.80% ±2.9650.94% ±4.4182.25% ±1.0589.38% ±0.6778.42% ±0.4770.18% ±1.042.82%3.85%5.09%cần sửap85_not_met –
48 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1 vinai/phobert-base-v20.068 m1encoderzephyrus 6 6 86M 19.6 ms 92.76% ±0.7586.00% ±1.1789.73% ±1.8586.79% ±1.2488.23% ±0.9193.20% ±0.6393.85% ±1.2955.47% ±4.9789.13% ±0.5991.30% ±0.5781.79% ±3.6081.85% ±3.3491.62% ±0.8782.96% ±2.6085.65% ±3.4482.69% ±2.4487.98% ±0.9292.12% ±0.7494.13% ±1.4155.07% ±5.0886.41% ±0.8091.11% ±0.5481.69% ±3.6281.43% ±3.522.18%2.95%4.46%cần sửap85_not_met –
49 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_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, nhãn m1 vinai/phobert-base-v20.052 m1encoderzephyrus 3 3 86M 20.9 ms 92.15% ±0.3384.37% ±1.3088.91% ±2.4685.40% ±1.1387.52% ±0.6692.72% ±0.2193.33% ±1.1852.32% ±4.1987.40% ±1.8889.70% ±0.2881.79% ±5.0071.12% ±4.7090.85% ±0.4880.03% ±2.0284.41% ±4.9080.57% ±2.7587.13% ±0.7391.51% ±0.3493.17% ±1.7151.65% ±4.5284.49% ±2.6089.42% ±0.2181.42% ±5.3370.79% ±4.332.82%2.05%5.01%okp85_not_met –
50 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-v20.058 m1encoderzephyrus 3 3 135M 19.6 ms 92.06% ±0.3981.79% ±1.2889.47% ±2.5084.45% ±1.6387.36% ±0.7392.56% ±0.3392.31% ±0.7753.92% ±2.7085.66% ±1.0489.42% ±0.5584.62% ±0.7768.50% ±1.7790.77% ±0.5277.42% ±2.1086.03% ±4.7480.17% ±2.9687.01% ±0.8691.38% ±0.4392.62% ±0.0053.26% ±3.1482.43% ±1.2789.14% ±0.5284.70% ±1.2567.87% ±1.633.08%2.82%4.62%okp85_not_met –
51 e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.056 m1decoderlegion 3 3 13M 49.4 ms 91.94% ±0.6383.12% ±1.4288.93% ±1.4384.71% ±0.1486.71% ±1.1792.58% ±0.5093.08% ±0.7750.62% ±3.7485.31% ±0.9089.84% ±0.3478.72% ±3.6475.92% ±3.4090.66% ±0.6979.25% ±1.5785.53% ±1.9980.82% ±0.2286.40% ±1.1191.41% ±0.5693.72% ±1.7150.05% ±3.4481.60% ±1.2089.69% ±0.4079.23% ±4.2175.63% ±3.212.82%3.08%5.34%okp85_not_met –
52 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-3-270m0.098 m1decoderzephyrus 3 3 4M 55.8 ms 90.77% ±0.6479.85% ±0.1287.69% ±2.4282.03% ±1.0885.10% ±1.3091.45% ±0.5690.77% ±1.3348.54% ±2.5384.13% ±0.5188.10% ±0.4478.21% ±1.6071.52% ±4.0289.27% ±0.9175.92% ±0.7984.64% ±3.6178.01% ±1.8984.77% ±1.4690.09% ±0.7891.26% ±1.2548.16% ±3.1880.28% ±0.9387.82% ±0.7177.87% ±2.4670.97% ±3.853.59%4.10%5.64%okp85_not_met –
53 e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.029 m1encoderlegion 3 3 304M 12.6 ms 92.79% ±0.2684.33% ±1.1388.33% ±1.9985.28% ±0.5288.08% ±0.5593.27% ±0.2493.08% ±1.5454.45% ±1.2986.04% ±0.2190.23% ±0.7782.82% ±0.4473.06% ±3.8891.69% ±0.2479.93% ±0.5983.60% ±3.5280.40% ±1.3787.82% ±0.4992.26% ±0.2493.72% ±1.2553.81% ±1.5682.30% ±0.4990.16% ±0.7283.88% ±1.2572.73% ±3.792.31%4.10%4.55%cần sửap85_not_met –
54 e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.034 m1encoderzephyrus 3 3 111M 19.7 ms 91.28% ±0.8382.27% ±2.3791.46% ±1.0285.14% ±1.8186.29% ±0.6892.09% ±0.6994.10% ±1.1847.86% ±2.3786.53% ±1.3190.27% ±0.5780.00% ±3.5381.56% ±4.7190.51% ±0.3579.53% ±2.3090.37% ±1.8382.83% ±2.0686.44% ±0.2791.31% ±0.3393.99% ±2.0649.72% ±0.5383.66% ±1.1690.14% ±0.5479.78% ±3.4181.80% ±4.022.56%2.31%6.00%okp85_not_met –
55 e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.051 m1encoderzephyrus 3 3 111M 18.3 ms 91.56% ±0.5681.43% ±2.7689.76% ±3.1583.95% ±3.2386.35% ±0.8692.26% ±0.4993.08% ±0.7749.95% ±0.9785.06% ±2.8389.96% ±0.6881.03% ±0.8976.34% ±1.0190.49% ±0.6478.34% ±3.0887.78% ±5.3181.02% ±4.1986.24% ±1.0191.27% ±0.5693.72% ±0.9550.45% ±0.7682.07% ±2.7889.77% ±0.7681.15% ±0.8275.98% ±1.282.31%3.08%5.45%okp85_not_met –
56 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.040 m1encoderzephyrus 3 3 111M 16.6 ms 91.31% ±1.1781.50% ±1.6587.07% ±1.3882.79% ±1.3785.97% ±2.1492.11% ±0.9794.10% ±0.8947.82% ±4.1586.19% ±1.6590.57% ±0.7182.82% ±1.9477.99% ±3.5689.80% ±1.2776.48% ±1.0082.08% ±1.2677.64% ±0.6285.51% ±2.0590.77% ±1.0493.99% ±0.4746.88% ±3.8382.67% ±2.0390.35% ±0.7782.79% ±1.6477.47% ±3.831.28%3.08%6.04%okp85_not_met –
57 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.057 m1encoderzephyrus 3 3 111M 19.6 ms 91.93% ±0.9182.27% ±1.5689.53% ±0.4184.47% ±1.1187.13% ±1.0792.57% ±0.7193.59% ±1.1851.27% ±4.5285.81% ±0.6989.98% ±0.1777.69% ±0.7778.52% ±1.7590.70% ±0.6978.73% ±1.6486.50% ±0.4380.92% ±1.2986.82% ±0.8491.44% ±0.5593.72% ±1.2550.96% ±3.3082.83% ±1.1589.73% ±0.2177.60% ±0.4778.05% ±1.781.79%3.08%5.25%okp85_not_met –
58 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-base0.070 m1encoderzephyrus 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 –
59 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1factorized · train thật, nhãn m1 jhu-clsp/mmBERT-base0.047 m1encoderzephyrus 3 3 111M 20.2 ms 91.87% ±0.3282.88% ±2.3587.77% ±0.8184.20% ±1.2986.75% ±0.7092.48% ±0.2291.28% ±0.8950.94% ±2.2685.84% ±1.5089.46% ±0.6878.21% ±3.1176.81% ±1.5990.54% ±0.4178.47% ±3.2983.37% ±2.1679.67% ±2.4486.38% ±0.6991.26% ±0.2991.53% ±0.9550.31% ±2.2881.97% ±2.3689.17% ±0.7277.87% ±3.2876.18% ±1.692.56%5.13%5.15%cần sửap85_not_met –
60 e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1gen_label · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 Qwen/Qwen3-1.7B-Base0.095 m1decoderlegion 3 3 17M 52.9 ms 93.03% ±0.7186.71% ±0.6389.92% ±2.8887.34% ±1.5488.53% ±1.0593.53% ±0.5991.28% ±1.9454.11% ±4.0188.01% ±0.2590.74% ±0.2582.05% ±2.9173.82% ±2.2491.76% ±0.8883.77% ±0.8787.29% ±3.6784.40% ±1.8788.07% ±1.1692.36% ±0.7490.98% ±1.6452.95% ±4.1784.87% ±0.2690.52% ±0.1682.24% ±2.8873.31% ±2.292.31%5.64%4.56%okp85_not_met –
61 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-base0.012 m1encoderzephyrus 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 –
62 e3_ensemble_seed_kd_a05_t1_m1ensemble trung bình nhân trên 1 phương pháp · nhãn m1 –0.078 m1ensemblelegion 1 1 N/A 21.2 ms 93.26%86.44%90.22%87.51%89.36%93.63%94.62%57.75%90.58%92.40%83.85%83.85%92.28%82.63%86.00%83.40%89.18%92.70%95.08%57.43%87.91%92.15%83.61%82.93%1.54%3.08%4.04%cần sửap85_not_met –
63 e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1gemma-3-1b-pt · 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 unsloth/gemma-3-1b-pt0.042 m1encoderlegion 3 3 100M 55.9 ms 91.63% ±0.4084.17% ±0.1588.57% ±1.6885.08% ±1.0286.24% ±0.5892.31% ±0.2993.85% ±2.0449.59% ±2.8487.09% ±0.5590.23% ±0.6276.92% ±4.0082.18% ±3.9390.41% ±0.3780.50% ±0.6085.14% ±3.2481.44% ±2.0086.00% ±0.5591.17% ±0.2994.26% ±2.1749.53% ±2.4183.49% ±0.6489.97% ±0.6076.50% ±4.0481.88% ±4.152.31%3.59%5.60%okp85_not_met –
64 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_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, nhãn m1 vinai/phobert-base-v20.024 m1encoderzephyrus 3 3 86M 19.5 ms 91.70% ±0.3984.75% ±0.7488.31% ±2.6185.13% ±1.2186.89% ±0.9492.37% ±0.3994.10% ±1.1849.35% ±1.0086.90% ±0.9190.34% ±0.6284.87% ±0.8973.07% ±0.7690.34% ±0.3980.47% ±1.2183.41% ±4.1280.19% ±2.5086.50% ±0.9391.11% ±0.4193.99% ±0.9548.60% ±0.8183.64% ±1.0690.10% ±0.6284.97% ±1.2572.66% ±0.742.82%2.05%5.64%okp85_not_met –
65 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_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 2e-05, 6 epoch · train thật, nhãn m1 vinai/phobert-base-v20.031 m1encoderzephyrus 3 3 86M 22.1 ms 92.01% ±1.0884.43% ±0.7288.05% ±2.8585.03% ±1.8886.90% ±1.7392.56% ±0.9092.05% ±2.2251.77% ±5.1286.39% ±1.4389.45% ±1.3580.77% ±0.7771.71% ±5.1490.73% ±1.2679.83% ±1.1883.70% ±5.1280.26% ±3.2786.54% ±1.7591.37% ±1.0692.35% ±2.3750.93% ±5.1083.20% ±1.6789.11% ±1.4280.33% ±0.8270.82% ±5.352.05%4.62%5.09%okp85_not_met –
66 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t2_tsa_m1 vinai/phobert-base-v20.070 m1encoderzephyrus 3 3 86M 26.8 ms 92.55% ±0.5185.08% ±1.1991.15% ±0.3587.16% ±0.5987.93% ±0.3493.00% ±0.4191.54% ±2.3154.68% ±3.3988.39% ±1.3691.05% ±0.8980.26% ±1.7880.05% ±1.9791.33% ±0.6582.11% ±1.5089.16% ±1.0584.46% ±0.8887.61% ±0.4491.87% ±0.5392.35% ±2.0654.21% ±3.7385.59% ±1.5091.11% ±0.8482.24% ±1.8980.47% ±1.102.82%4.10%4.46%okp85_not_met –
67 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_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-v20.031 m1encoderzephyrus 3 3 135M 22.5 ms 92.25% ±0.7584.70% ±0.9588.47% ±1.8585.50% ±1.4187.20% ±0.9492.77% ±0.5894.10% ±0.4453.44% ±4.4786.11% ±0.6289.52% ±0.6083.33% ±2.3571.92% ±3.5190.99% ±0.9480.44% ±1.7884.10% ±3.3481.04% ±2.6786.84% ±1.0491.60% ±0.7594.26% ±0.8252.69% ±4.6182.77% ±0.7889.20% ±0.6883.06% ±2.5071.23% ±3.702.31%2.56%4.83%okp85_not_met –
68 e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 5e-05, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.096 m1decoderlegion 3 3 13M 49.3 ms 92.01% ±0.9879.31% ±2.9985.03% ±4.9281.15% ±3.9287.39% ±1.1992.56% ±0.7590.77% ±0.0053.06% ±6.0684.44% ±0.8490.11% ±0.2782.05% ±1.9473.46% ±2.6690.80% ±1.2675.58% ±2.2181.95% ±2.0777.58% ±2.3987.10% ±1.2891.45% ±0.9690.71% ±0.4753.09% ±6.5780.09% ±0.9289.94% ±0.1782.51% ±1.2573.55% ±2.595.13%3.08%4.77%okp85_not_met –
69 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.038 m1encoderlegion 3 3 304M 12.9 ms 91.69% ±0.6782.97% ±0.6289.23% ±1.1384.79% ±0.1986.68% ±0.3692.36% ±0.5294.10% ±1.1849.81% ±3.2586.30% ±0.6990.46% ±0.1483.85% ±2.6676.60% ±1.5890.30% ±0.7578.90% ±0.2285.84% ±1.9680.85% ±0.7486.32% ±0.4191.11% ±0.6095.08% ±0.8249.08% ±3.0482.39% ±0.5990.51% ±0.2686.07% ±2.1776.48% ±1.742.31%2.56%5.57%okp85_not_met –
70 e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1factorized · train thật + synth ×1, nhãn m1 jhu-clsp/mmBERT-base0.061 m1encoderzephyrus 3 3 111M 19.5 ms 91.08% ±0.4182.19% ±2.6588.90% ±4.5083.72% ±1.7486.09% ±0.7291.91% ±0.3793.85% ±2.0447.34% ±1.9786.20% ±1.1489.55% ±0.2683.59% ±4.2471.68% ±3.1589.96% ±0.6578.07% ±3.4485.48% ±8.0779.65% ±4.5685.99% ±0.8290.85% ±0.5894.26% ±2.1747.88% ±2.7282.92% ±1.8189.34% ±0.2383.88% ±4.5171.33% ±3.932.31%3.59%6.10%oktest_informedp85_not_met –
71 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.017 m1encoderzephyrus 3 3 86M 14.2 ms 92.44% ±0.2786.49% ±2.2985.90% ±0.3884.72% ±0.9387.64% ±0.5692.92% ±0.2292.56% ±1.7853.52% ±1.5885.15% ±1.5189.93% ±0.7983.85% ±2.0471.61% ±3.1691.21% ±0.3581.68% ±6.1778.82% ±0.4777.00% ±0.9087.35% ±0.6391.73% ±0.2893.17% ±1.7153.00% ±1.8480.27% ±2.0389.75% ±0.8484.70% ±1.8971.00% ±3.302.56%3.59%4.70%okp85_not_met –
72 e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.026 m1encoderzephyrus 3 3 135M 9.2 ms 92.83% ±0.7285.47% ±0.5087.24% ±1.6685.49% ±1.4688.06% ±1.3593.20% ±0.5790.77% ±1.3357.55% ±6.0586.01% ±0.3790.03% ±0.4682.82% ±4.2473.36% ±2.5691.74% ±0.7580.16% ±2.4681.66% ±3.0579.69% ±3.0287.85% ±1.2992.15% ±0.5991.80% ±1.4257.07% ±5.6982.11% ±0.9489.87% ±0.5183.61% ±4.9272.81% ±2.643.59%3.85%3.98%okp85_not_met –
73 e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_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 MLP 512 · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.065 m1encoderzephyrus 3 3 87M 20.9 ms 92.08% ±0.3282.97% ±1.6688.83% ±1.2584.81% ±0.9787.17% ±0.6492.61% ±0.2492.56% ±1.1852.98% ±2.8185.54% ±0.6389.26% ±0.3983.33% ±1.9469.34% ±1.6090.80% ±0.4778.76% ±1.9784.84% ±2.1580.53% ±1.3786.78% ±0.8191.42% ±0.3392.62% ±0.8252.41% ±3.4781.74% ±0.4588.92% ±0.3783.06% ±2.0668.51% ±1.483.08%2.31%4.82%okp85_not_met –
74 e2_xlmr_base_cross_logitadjust_real_frozenemb_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE điều chỉnh logit theo prior (τ=1) · lr 2e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.018 m1encoderzephyrus 3 3 86M 10.6 ms 90.80% ±0.5580.45% ±0.9188.30% ±0.4882.89% ±0.7285.62% ±0.8791.58% ±0.4393.59% ±1.1847.18% ±2.5085.03% ±0.4088.52% ±0.6083.33% ±3.2067.84% ±2.6189.17% ±0.6875.70% ±0.8883.78% ±0.6878.09% ±0.8285.13% ±0.9890.11% ±0.5493.99% ±1.8946.40% ±2.8081.87% ±0.8188.25% ±0.6184.43% ±3.5767.43% ±2.212.31%3.33%6.13%okp85_not_met –
75 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.026 m1encoderlegion 3 3 304M 13.1 ms 93.19% ±0.7883.87% ±0.7190.44% ±0.5286.24% ±0.8888.85% ±0.9793.60% ±0.5992.05% ±1.6058.24% ±6.4786.40% ±0.1590.78% ±0.1881.54% ±1.5477.57% ±0.5592.00% ±0.9379.83% ±0.6887.39% ±0.5182.41% ±0.8988.45% ±0.9692.51% ±0.6991.80% ±2.1757.20% ±6.3682.51% ±0.1990.59% ±0.2381.69% ±1.7177.07% ±0.612.56%4.10%3.93%okp85_not_met –
76 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1factorized · train thật, nhãn m1 jhu-clsp/mmBERT-base0.072 m1encoderzephyrus 3 3 111M 22.0 ms 92.32% ±0.4382.68% ±2.1187.19% ±1.4684.02% ±1.2286.95% ±0.4392.78% ±0.2989.23% ±2.6654.26% ±3.5685.72% ±1.5889.18% ±1.0778.46% ±2.6677.06% ±1.2691.01% ±0.4777.93% ±2.8382.78% ±2.4379.33% ±2.3286.53% ±0.4391.57% ±0.3189.34% ±2.8453.28% ±3.2981.87% ±2.4088.89% ±1.1278.14% ±2.6376.46% ±1.303.59%5.90%4.43%okp85_not_met –
77 e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 3e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.082 m1encoderzephyrus 3 3 111M 20.5 ms 92.17% ±0.1882.26% ±1.2087.90% ±1.6083.85% ±1.1786.97% ±0.6492.70% ±0.2393.08% ±1.5453.05% ±1.7585.71% ±1.6089.25% ±0.9482.05% ±2.9172.44% ±3.5390.83% ±0.3277.74% ±1.9483.96% ±2.9879.37% ±2.1286.58% ±0.7691.48% ±0.3593.44% ±1.6452.36% ±1.9081.93% ±2.5288.95% ±0.9981.97% ±2.9671.83% ±3.772.31%2.82%4.82%okp85_not_met –
78 e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1Qwen3-1.7B Base · 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 Qwen/Qwen3-1.7B-Base0.067 m1encoderlegion 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 –
79 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/visobert0.037 m1encoderzephyrus 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 –
80 e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.027 m1encoderzephyrus 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 –
81 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-base0.057 m1encoderzephyrus 3 3 110M 36.2 ms 91.96% ±1.0086.71% ±1.2986.66% ±1.4085.18% ±0.8087.47% ±1.3392.62% ±0.7893.59% ±1.9450.41% ±5.0486.52% ±1.4390.33% ±1.0877.44% ±3.1175.63% ±5.5990.80% ±1.0081.80% ±1.2780.90% ±2.6279.32% ±1.9387.22% ±1.2791.52% ±0.8093.72% ±1.7150.15% ±4.2682.41% ±1.8290.10% ±1.1677.32% ±2.8875.01% ±5.942.56%3.59%5.46%okp85_not_met –
82 e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_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 2e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 vinai/phobert-base-v20.206 m1encoderzephyrus 3 3 86M 18.8 ms 93.33% ±1.3084.75% ±0.8290.66% ±1.4586.95% ±0.4689.28% ±1.8293.57% ±1.0188.72% ±5.0066.36% ±13.3786.51% ±0.5690.05% ±0.5381.28% ±3.2074.93% ±1.6392.22% ±1.5581.13% ±1.8088.17% ±2.9483.55% ±1.6488.94% ±1.9492.53% ±1.2288.80% ±4.9465.57% ±13.5882.72% ±1.2989.75% ±0.5681.15% ±3.7674.23% ±1.764.36%5.38%2.93%okp85_not_met –
83 e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_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 cổng trộn u/v, head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.034 m1encoderzephyrus 3 3 87M 20.6 ms 92.76% ±0.3084.22% ±1.8189.03% ±2.0785.76% ±0.7088.15% ±0.5193.12% ±0.2190.51% ±1.1857.48% ±3.5887.11% ±1.4590.25% ±0.4783.85% ±1.3375.13% ±3.4291.50% ±0.4679.74% ±2.4084.55% ±4.4180.88% ±2.1587.70% ±0.7091.92% ±0.3690.71% ±1.2556.77% ±3.9083.79% ±1.6689.85% ±0.4883.61% ±1.4274.60% ±3.543.08%4.87%3.93%okp85_not_met –
84 e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_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 2e-05, 4 epoch · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1 vinai/phobert-base-v20.052 m1encoderzephyrus 3 3 86M 19.4 ms 92.59% ±0.9584.31% ±1.1986.73% ±0.5984.57% ±0.6787.68% ±1.6992.99% ±0.7191.54% ±4.2856.25% ±7.0686.41% ±0.5389.74% ±0.6879.74% ±1.1875.30% ±0.6991.31% ±1.1078.82% ±1.4480.93% ±0.6378.76% ±0.6887.27% ±1.7091.79% ±0.8491.53% ±4.6655.26% ±6.7582.73% ±1.0589.41% ±0.6879.51% ±1.4274.62% ±0.673.59%3.59%4.29%oktest_informedp85_not_met –
85 e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_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 + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1 vinai/phobert-base-v20.033 m1encoderzephyrus 3 3 135M 20.4 ms 92.42% ±0.7084.62% ±1.4486.60% ±1.1384.59% ±0.7987.37% ±1.3792.83% ±0.5592.56% ±2.4755.37% ±4.4986.71% ±0.5989.62% ±0.7580.26% ±1.1876.34% ±1.1091.25% ±0.7879.32% ±3.3680.86% ±2.4178.83% ±2.2787.09% ±1.3391.70% ±0.6292.62% ±2.8455.01% ±4.3483.26% ±1.0489.24% ±0.8379.78% ±1.2575.46% ±1.483.33%3.08%4.41%oktest_informedp85_not_met –
86 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.144 m1encoderzephyrus 3 3 111M 18.3 ms 91.86% ±0.5382.22% ±0.3287.69% ±1.5283.85% ±0.7886.80% ±0.5892.47% ±0.3891.54% ±0.7751.50% ±3.7286.11% ±1.7689.41% ±1.4282.56% ±4.4471.08% ±8.7690.42% ±0.6177.30% ±0.6083.22% ±2.8378.96% ±1.7386.33% ±0.6491.17% ±0.4491.53% ±1.2550.29% ±3.6481.91% ±2.9489.13% ±1.5682.51% ±4.5170.28% ±9.092.05%4.62%5.07%okp85_not_met –
87 e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.082 m1encoderzephyrus 3 3 111M 18.6 ms 91.86% ±0.9981.74% ±1.9491.71% ±0.5484.98% ±1.2086.76% ±1.2992.48% ±0.7191.03% ±4.6451.91% ±7.2285.25% ±0.4389.34% ±0.5279.74% ±2.9173.54% ±3.3490.56% ±1.1378.72% ±2.4090.96% ±1.2782.24% ±1.9286.38% ±1.2991.31% ±0.8290.98% ±4.5651.25% ±7.0082.10% ±0.4689.09% ±0.5279.78% ±3.4172.84% ±3.402.82%3.33%5.09%cần sửap85_not_met –
88 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.108 m1encoderzephyrus 3 3 111M 18.7 ms 92.83% ±0.5383.59% ±1.7088.90% ±0.3885.37% ±1.1887.93% ±0.7193.21% ±0.3591.79% ±2.3557.88% ±5.4486.89% ±1.0790.23% ±0.3684.36% ±3.4778.83% ±2.5091.64% ±0.5679.35% ±1.8985.53% ±1.1781.35% ±1.5687.62% ±0.6592.11% ±0.3692.08% ±2.3757.03% ±5.2383.31% ±0.9790.00% ±0.3484.43% ±3.2878.36% ±2.682.82%4.87%3.96%okp85_not_met –
89 e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-0.6B-Base0.075 m1decoderlegion 3 3 10M 45.9 ms 91.82% ±1.2682.16% ±1.1090.95% ±1.8585.05% ±1.5887.47% ±2.0392.51% ±0.9493.59% ±1.6050.42% ±6.5586.76% ±0.8590.91% ±0.8181.28% ±3.5577.60% ±2.1790.44% ±1.5478.56% ±1.4688.30% ±2.6781.55% ±1.9387.05% ±2.1491.26% ±1.1693.44% ±1.6449.73% ±6.7483.70% ±1.0690.74% ±0.7781.42% ±3.3177.49% ±2.081.54%3.85%5.49%okp85_not_met –
90 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t1_tsa_m1 vinai/phobert-base-v20.070 m1encoderzephyrus 3 3 86M 22.7 ms 92.30% ±0.3684.72% ±1.5390.64% ±2.0086.54% ±1.0387.53% ±0.3492.80% ±0.2693.59% ±1.7853.51% ±3.0288.15% ±1.4991.16% ±0.6082.31% ±1.5481.75% ±2.4190.99% ±0.4581.43% ±1.8688.09% ±3.7083.41% ±2.0987.16% ±0.4191.58% ±0.3593.72% ±1.7152.86% ±3.2685.04% ±1.7791.15% ±0.6283.88% ±1.2581.73% ±2.562.31%2.56%4.77%okp85_not_met –
91 e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1Qwen3-0.6B Base · 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 Qwen/Qwen3-0.6B-Base0.077 m1encoderlegion 3 3 97M 50.7 ms 90.25% ±1.1883.16% ±2.0887.92% ±1.6584.16% ±2.3285.86% ±1.1191.14% ±0.8590.26% ±3.8745.47% ±5.0386.61% ±1.6189.26% ±0.0875.13% ±1.9473.06% ±1.4389.19% ±1.9279.29% ±3.8684.02% ±3.4580.31% ±4.2385.90% ±1.6590.09% ±1.4290.44% ±4.1347.40% ±7.5383.31% ±2.2588.91% ±0.1574.86% ±2.5072.48% ±1.873.08%6.15%6.45%cần sửap85_not_met –
92 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_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-v20.093 m1encoderzephyrus 3 3 86M 20.4 ms 92.66% ±0.2683.33% ±2.3791.46% ±2.2186.17% ±0.4887.95% ±0.2193.11% ±0.2592.56% ±0.4455.82% ±1.3087.25% ±1.2290.33% ±0.7184.10% ±4.2476.45% ±5.0791.45% ±0.3179.83% ±1.9789.37% ±4.2482.81% ±0.5587.57% ±0.2592.00% ±0.3192.62% ±0.8254.88% ±1.3684.48% ±1.4390.09% ±0.6883.88% ±4.5175.97% ±5.022.82%3.08%4.28%okp85_not_met –
93 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_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, nhãn m1 vinai/phobert-base-v20.104 m1encoderzephyrus 3 3 86M 22.0 ms 92.78% ±1.1685.85% ±1.4588.85% ±2.3486.35% ±1.5688.02% ±2.1393.14% ±0.8891.54% ±3.0858.93% ±8.8788.51% ±0.8890.56% ±0.3985.13% ±1.6077.70% ±3.7691.56% ±1.3483.19% ±3.2584.84% ±3.9682.19% ±2.6087.64% ±2.1691.97% ±1.0390.98% ±3.2858.33% ±8.9585.46% ±0.9490.30% ±0.3884.70% ±1.2577.61% ±3.433.59%3.59%3.90%okp85_not_met –
94 e2_xlmr_base_cross_ce_real_frozenemb_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.044 m1encoderzephyrus 3 3 86M 12.1 ms 91.22% ±1.2981.76% ±0.9487.01% ±2.0683.20% ±1.9886.03% ±1.6391.87% ±0.9990.26% ±3.5550.47% ±7.9784.45% ±1.9888.35% ±1.2780.51% ±5.8868.70% ±10.7889.80% ±1.6275.96% ±2.8281.80% ±4.3577.52% ±3.9785.71% ±1.9090.56% ±1.2790.98% ±3.5749.96% ±8.4280.26% ±3.4088.21% ±1.5281.15% ±5.3768.27% ±11.223.59%5.13%5.37%okp85_not_met –
95 e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE trọng số căn nghịch tần suất · lr 2e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.019 m1encoderzephyrus 3 3 86M 10.9 ms 91.36% ±0.7981.57% ±1.0089.63% ±1.1484.31% ±1.0686.38% ±0.9991.95% ±0.5292.56% ±3.2051.70% ±5.3984.79% ±0.2489.11% ±0.5381.28% ±3.8772.56% ±6.8989.84% ±0.9977.46% ±1.0186.63% ±2.6180.41% ±1.5785.92% ±1.1190.55% ±0.6693.44% ±2.4650.80% ±5.5580.83% ±0.4288.92% ±0.6682.24% ±3.4172.17% ±7.061.79%4.62%5.15%okp85_not_met –
96 e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_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 điều chỉnh logit theo prior (τ=0.25) · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.049 m1encoderzephyrus 3 3 86M 18.9 ms 91.90% ±0.2481.58% ±1.1788.85% ±1.5283.91% ±0.6487.05% ±0.3692.47% ±0.2993.08% ±1.3351.80% ±0.8185.27% ±0.5389.72% ±0.3283.59% ±3.1171.55% ±1.5390.63% ±0.2977.00% ±0.6884.98% ±3.0179.40% ±1.2786.76% ±0.3891.30% ±0.3393.17% ±1.2551.38% ±1.1881.82% ±0.5989.45% ±0.2583.61% ±3.5771.07% ±1.623.08%1.79%5.06%okp85_not_met –
97 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_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 1e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.159 m1encoderzephyrus 3 3 86M 22.1 ms 92.20% ±1.0382.98% ±2.7185.85% ±5.2083.15% ±4.9187.48% ±1.3592.58% ±0.7991.03% ±1.6056.24% ±7.9986.66% ±0.8789.21% ±0.2484.87% ±3.7969.61% ±1.5491.06% ±1.0979.63% ±2.1383.85% ±2.5380.71% ±2.5687.22% ±1.2691.52% ±0.8191.53% ±1.8955.95% ±7.3783.22% ±1.4189.00% ±0.1985.52% ±4.0469.16% ±1.463.85%4.10%4.28%okp85_not_met –
98 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsa_m1 vinai/phobert-base-v20.085 m1encoderzephyrus 3 3 86M 22.9 ms 92.68% ±0.8583.79% ±0.9189.89% ±2.5085.89% ±1.8488.07% ±1.3593.09% ±0.6892.05% ±2.3556.31% ±5.0987.91% ±0.8490.98% ±0.7881.54% ±2.7780.91% ±2.1391.43% ±0.9979.42% ±1.8786.44% ±4.6981.76% ±3.3187.71% ±1.3991.92% ±0.8092.08% ±2.0655.62% ±5.0084.66% ±0.9690.67% ±0.8681.15% ±2.9680.27% ±2.632.05%3.59%4.23%okp85_not_met –
99 e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1 Qwen/Qwen3-Embedding-4B0.061 m1decoderkaggle 2 2 79M 126.5 ms 90.03% ±0.2482.18% ±4.2785.99% ±2.3882.11% ±1.7184.23% ±0.8890.71% ±0.3690.00% ±3.2647.85% ±1.0485.27% ±1.8487.00% ±2.0669.62% ±13.6076.22% ±5.9388.16% ±0.2978.88% ±4.6082.59% ±4.4178.03% ±1.3983.47% ±0.9289.00% ±0.5189.34% ±3.4846.58% ±1.5382.67% ±2.7086.43% ±2.2668.85% ±13.9175.56% ±6.293.46%5.00%5.72%okp85_not_met –
100 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.025 m1encoderzephyrus 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 –
101 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1Qwen3-Embedding-0.6B · cross-attention (Q = câu KH, K/V = câu bot), backbone đóng băng · CE trọng số căn nghịch tần suất, label smoothing 0.05 · lr 0.0003, 40 epoch · train thật, nhãn m1 Qwen/Qwen3-Embedding-0.6B0.152 m1decoderzephyrus 3 3 13M 62.5 ms 90.77% ±0.8780.02% ±0.7486.96% ±0.6182.32% ±0.4984.83% ±1.4191.24% ±0.7684.10% ±0.4452.38% ±5.0383.89% ±1.0086.45% ±0.4673.59% ±3.1168.33% ±1.4589.05% ±1.0875.94% ±0.5384.13% ±0.9778.33% ±0.1784.15% ±1.5489.65% ±0.9583.33% ±0.4751.01% ±5.3380.40% ±1.6985.89% ±0.4972.68% ±3.3167.34% ±1.847.44%5.38%4.52%okp85_not_met –
102 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_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 5e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.035 m1encoderzephyrus 3 3 86M 20.7 ms 92.20% ±1.0083.17% ±0.7289.42% ±0.8385.20% ±0.7587.48% ±1.0892.73% ±0.7793.59% ±1.6053.58% ±5.4287.16% ±0.1990.25% ±0.9782.05% ±4.2475.15% ±4.3390.89% ±1.1678.83% ±0.5285.47% ±1.6780.89% ±0.9587.07% ±1.1491.53% ±0.9093.72% ±1.8952.68% ±5.4384.03% ±0.1589.93% ±1.0381.69% ±4.5174.42% ±4.532.31%2.05%4.82%cần sửap85_not_met –
103 e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-base0.052 m1encoderzephyrus 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 –
104 e2_visobert_cross_ce_real_frozenemb_lr2e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 uitnlp/visobert0.027 m1encoderzephyrus 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 –
105 e2_visobert_cross_ce_real_frozenemb_lr1e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 uitnlp/visobert0.186 m1encoderzephyrus 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 –
106 e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.161 m1encoderzephyrus 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 –
107 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-0.6B0.149 m1tfidf_hybridzephyrus 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 –
108 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-IDF0.207 m1tfidfzephyrus 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 –
109 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-IDF0.164 m1tfidfzephyrus 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 –
110 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-IDF0.280 m1tfidfzephyrus 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 –
111 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-IDF0.178 m1tfidfzephyrus 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 –
112 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-IDF0.267 m1tfidfzephyrus 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 –
113 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-IDF0.298 m1tfidfzephyrus 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 –
114 e2_dual_qwen3emb06b_concat_ce_real_frozen_m1Qwen3-Embedding-0.6B · dual encoder (câu bot và câu KH mã hoá riêng, pooling last_token), backbone đóng băng · fusion [u; v; |u−v|; u·v], head MLP 512 · CE · lr head 0.001, 30 epoch · train thật, nhãn m1 Qwen/Qwen3-Embedding-0.6B0.227 m1decoderzephyrus 3 3 2M 54.5 ms 88.18% ±0.7683.85% ±1.9380.61% ±2.4680.49% ±1.0681.14% ±1.0388.80% ±0.6185.90% ±1.1845.40% ±3.3082.63% ±1.3583.28% ±1.0779.49% ±0.4459.16% ±3.0986.17% ±0.8381.25% ±3.6974.50% ±3.6774.43% ±2.5080.44% ±1.0286.86% ±0.6585.52% ±0.9544.57% ±3.2177.54% ±2.3082.75% ±1.1979.78% ±0.4758.51% ±3.326.41%6.41%6.07%okp85_not_met –
115 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-IDF0.282 m1tfidfzephyrus 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 –
116 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-IDF0.190 m1tfidfzephyrus 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 –
117 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-IDF0.310 m1tfidfzephyrus 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 –
118 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-IDF0.356 m1tfidfzephyrus 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 –– m1encoder 0 1 lỗi 0 N/A N/A N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/Afailed –
– e2_bamibert_cross_ce_real_frozenemb_lr2e5_m1 –– m1encoder 0 1 lỗi 0 N/A N/A N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/Afailed –
– e2_bamibert_cross_ce_real_frozenemb_lr4e5_m1 –– m1encoder 0 1 lỗi 0 N/A N/A N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/Afailed –
– e3_fbverifier_devtop1_familygeomean_fbf1_m1fallback_verifier · train thật, nhãn m1 –hard m1ensemblezephyrus 1 1 N/A 271.8 ms 95.59%87.45%92.64%89.62%92.37%95.61%N/A83.46%89.62%92.37%85.38%83.46%95.01%84.11%91.36%86.77%92.36%95.06%N/A83.33%86.77%92.36%86.07%83.33%5.38%6.15%0.99%ok –
– e3_fbverifier_devtop1_familygeomean_macrof1_m1fallback_verifier · train thật, nhãn m1 –hard m1ensemblezephyrus 1 1 N/A 271.8 ms 95.59%87.45%92.64%89.62%92.37%95.61%N/A83.46%89.62%92.37%85.38%83.46%95.01%84.11%91.36%86.77%92.36%95.06%N/A83.33%86.77%92.36%86.07%83.33%5.38%6.15%0.99%ok –
1 e3_ensemble_mean_allensemble trung bình trên 23 phương pháp · nhãn v1 –0.065 v1ensemblezephyrus 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_allensemble chọn tham lam Caruana trên 23 phương pháp · nhãn v1 –0.070 v1ensemblezephyrus 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_lr4e5XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.023 v1encoderzephyrus 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-v20.011 v1encoderzephyrus 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_frozenembXLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE trọng số căn nghịch tần suất · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.021 v1encoderzephyrus 3 3 86M 16.1 ms 90.38% ±0.2778.15% ±2.9183.03% ±3.0779.82% ±2.9385.17% ±0.7190.85% ±0.2184.35% ±1.5653.17% ±2.2380.65% ±4.1086.53% ±0.8175.17% ±2.5767.47% ±2.7788.67% ±0.1675.95% ±3.4281.46% ±4.0677.76% ±3.6384.72% ±0.4989.27% ±0.1586.21% ±1.9951.41% ±1.5978.99% ±4.9186.31% ±0.7776.63% ±1.7666.30% ±2.331.02%12.93%4.79%okp85_not_met –
6 e3_prior_em_mmbert_bottypeEM điều chỉnh prior trên e2_mmbert_base_cross_ce_real_frozenemb_lr2e5, phạm vi theo loại câu bot (transductive) · nhãn v1 –0.082 v1post_hoc_transductivezephyrus 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_globalEM điều chỉnh prior trên e2_mmbert_base_cross_ce_real_frozenemb_lr2e5, phạm vi toàn bộ (transductive) · nhãn v1 –0.082 v1post_hoc_transductivezephyrus 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_lr2e5mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 jhu-clsp/mmBERT-base0.061 v1encoderzephyrus 3 3 111M 24.8 ms 90.46% ±0.5981.26% ±0.4887.10% ±1.8883.07% ±0.8884.84% ±1.0291.01% ±0.4985.71% ±1.0250.93% ±3.2884.38% ±1.1586.92% ±0.6974.49% ±4.0869.46% ±2.7688.87% ±0.7879.76% ±0.6586.83% ±2.2781.88% ±1.0884.48% ±1.0989.59% ±0.6488.51% ±1.1549.18% ±3.4283.57% ±1.2186.89% ±0.7477.01% ±4.6068.30% ±2.871.70%10.88%5.34%cần sửap85_not_met –
9 e2_xlmr_base_cross_ce_real_frozenembXLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.035 v1encoderzephyrus 3 3 86M 10.8 ms 90.83% ±0.0979.78% ±3.4582.64% ±0.7080.54% ±1.8185.49% ±0.3391.16% ±0.0682.99% ±3.5856.89% ±0.5880.28% ±1.5486.04% ±0.5671.43% ±1.0268.05% ±2.7189.28% ±0.1977.73% ±3.5681.12% ±0.7278.61% ±1.9185.13% ±0.3889.70% ±0.1584.67% ±4.0455.39% ±0.5878.65% ±1.7785.85% ±0.4273.18% ±1.7666.88% ±2.462.72%12.24%4.05%cần sửap85_not_met high
10 e2_xlmr_base_cross_logitadjust_real_frozenembXLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE điều chỉnh logit theo prior (τ=1) · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.035 v1encoderzephyrus 3 3 86M 14.7 ms 90.23% ±0.6876.57% ±1.4084.10% ±2.5078.73% ±1.6085.21% ±1.2990.84% ±0.6285.37% ±4.1250.92% ±2.6079.20% ±0.4687.02% ±0.8576.19% ±1.5665.34% ±1.9488.49% ±0.7374.71% ±1.4482.83% ±2.8476.87% ±1.6884.74% ±1.1989.28% ±0.6787.36% ±3.0449.04% ±2.1077.80% ±0.5686.83% ±0.6678.16% ±2.3063.79% ±1.351.70%11.22%5.30%okp85_not_met –
11 e2_qwen3_06b_base_cross_ce_real_loraQwen3-0.6B Base · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn v1 Qwen/Qwen3-0.6B-Base0.098 v1decoderlegion 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_allensemble stacking LR trên 23 phương pháp · nhãn v1 –0.088 v1ensemblezephyrus 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_loraQwen3-1.7B Base · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật, nhãn v1 Qwen/Qwen3-1.7B-Base0.083 v1decoderlegion 3 3 17M 53.1 ms 90.19% ±0.2180.63% ±1.3486.84% ±4.0082.43% ±2.0885.03% ±0.3390.76% ±0.0983.67% ±2.0450.23% ±2.7283.16% ±1.9587.57% ±0.7474.15% ±2.5770.98% ±4.1588.56% ±0.2879.00% ±1.8086.45% ±4.9081.10% ±2.6284.66% ±0.4089.29% ±0.1285.82% ±1.7648.72% ±2.7981.94% ±2.5187.45% ±0.9075.86% ±3.0469.94% ±4.531.70%13.27%5.36%okp85_not_met –
14 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 vinai/phobert-base-v20.022 v1encoderzephyrus 3 3 86M 12.6 ms 90.85% ±0.2683.85% ±2.7580.68% ±1.2681.05% ±1.8885.59% ±0.6491.20% ±0.2684.35% ±1.5655.00% ±0.9780.14% ±0.8387.11% ±0.8673.81% ±0.5969.39% ±2.4089.28% ±0.3981.46% ±4.2978.67% ±1.7578.42% ±2.6585.22% ±0.7089.73% ±0.3786.59% ±1.3353.32% ±1.2078.33% ±1.3787.02% ±0.9476.25% ±0.6668.21% ±2.531.70%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-base0.088 v1encoderzephyrus 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_lr4e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 intfloat/multilingual-e5-base0.028 v1encoderzephyrus 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_frozenembXLM-R base · chỉ câu KH · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.076 v1encoderzephyrus 3 3 86M 8.5 ms 90.38% ±0.7481.00% ±1.2081.57% ±2.9680.52% ±2.2585.48% ±0.7090.70% ±0.6282.99% ±3.1256.10% ±4.0981.36% ±1.6086.88% ±0.8076.19% ±2.3668.54% ±3.2588.69% ±0.9278.14% ±2.0579.83% ±3.8578.01% ±3.2185.04% ±0.7789.12% ±0.7985.44% ±3.5154.13% ±3.9679.41% ±2.2586.66% ±0.9278.16% ±3.0467.14% ±3.541.02%15.65%4.22%okp85_not_met –
18 e2_me5_base_cross_ce_real_frozenemb_lr2e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 intfloat/multilingual-e5-base0.056 v1encoderzephyrus 3 3 86M 10.9 ms 90.71% ±0.4079.46% ±3.5180.85% ±2.0279.53% ±2.6685.36% ±0.6291.10% ±0.3484.35% ±2.1255.12% ±1.6178.41% ±1.3085.67% ±0.8073.47% ±4.6865.15% ±1.5989.08% ±0.4476.65% ±4.1278.71% ±2.5176.89% ±3.1684.88% ±0.6789.60% ±0.3785.82% ±2.3952.97% ±1.6776.39% ±1.4785.47% ±0.7775.48% ±4.7963.64% ±1.783.40%9.86%4.42%cần sửap85_not_met –
19 e2_gemma3_270m_cross_ce_real_loragemma-3-270m · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn v1 unsloth/gemma-3-270m0.116 v1decoderzephyrus 3 3 4M 54.2 ms 89.57% ±0.6176.70% ±0.6982.66% ±1.6378.30% ±1.0683.66% ±0.4890.19% ±0.4285.03% ±2.5750.12% ±3.3779.71% ±1.6685.49% ±0.6572.45% ±1.0267.41% ±1.1187.77% ±0.7374.92% ±0.9281.48% ±2.1576.54% ±1.3483.32% ±0.6888.55% ±0.5388.12% ±1.7648.70% ±3.4077.47% ±1.0485.32% ±0.5574.33% ±1.3365.98% ±0.683.06%11.22%5.49%okp85_not_met –
20 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozenQwen3-Embedding-0.6B · cross-attention (Q = câu KH, K/V = câu bot), backbone đóng băng · CE trọng số căn nghịch tần suất, label smoothing 0.05 · lr 0.0003, 40 epoch · train thật, nhãn v1 Qwen/Qwen3-Embedding-0.6B0.114 v1decoderzephyrus 3 3 13M 82.5 ms 89.10% ±0.3177.88% ±1.2284.69% ±0.2979.96% ±1.0283.51% ±0.4689.67% ±0.2480.95% ±3.8649.00% ±1.1882.68% ±0.9084.79% ±0.6268.71% ±2.3665.60% ±1.8587.18% ±0.5076.11% ±1.2483.81% ±0.3278.35% ±1.1882.94% ±0.5287.91% ±0.3982.38% ±3.6947.21% ±1.4981.70% ±1.0084.54% ±0.6470.50% ±2.6564.34% ±1.606.46%9.86%5.43%okp85_not_met –
21 e2_mmbert_base_cross_ce_real_frozenemb_lr4e5mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 jhu-clsp/mmBERT-base0.075 v1encoderzephyrus 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_lr1e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 vinai/phobert-base-v20.026 v1encoderzephyrus 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-base0.123 v1encoderzephyrus 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_lr1e5mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 jhu-clsp/mmBERT-base0.120 v1encoderzephyrus 2 2 111M 23.9 ms 90.47% ±0.2276.81% ±1.5882.68% ±0.8078.75% ±0.6885.50% ±0.5990.88% ±0.2380.61% ±2.8955.08% ±1.0179.08% ±0.9186.16% ±0.8768.37% ±1.4467.46% ±4.7889.10% ±0.4375.07% ±1.1481.45% ±0.8277.08% ±0.4185.42% ±0.8189.64% ±0.4383.91% ±3.2553.93% ±1.5777.64% ±1.3686.09% ±1.0570.11% ±1.6366.08% ±5.062.55%13.78%4.24%cần sửap85_not_met –
25 e2_visobert_cross_ce_real_frozenemb_lr2e5ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 uitnlp/visobert0.093 v1encoderzephyrus 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_lr4e5ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 uitnlp/visobert0.066 v1encoderzephyrus 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-IDF0.123 v1tfidfzephyrus 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-IDF0.256 v1tfidfzephyrus 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-IDF0.125 v1tfidfzephyrus 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_lr1e5ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 uitnlp/visobert0.232 v1encoderzephyrus 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-IDF0.234 v1tfidfzephyrus 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 –– v1encoder 0 1 lỗi 0 N/A N/A N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/Afailed –
– e2_mdeberta_cross_ce_real_frozenemb_lr2e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 –– v1encoder 0 1 lỗi 0 N/A N/A N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/Afailed –
– e2_mdeberta_cross_ce_real_frozenemb_lr2e5_fp32mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 –– v1encoder 0 1 lỗi 0 N/A N/A N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/Afailed –
– e2_mdeberta_cross_ce_real_frozenemb_lr4e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 microsoft/mdeberta-v3-baseargmax v1encoderzephyrus 1 1 86M 29.1 ms 23.81%3.97%16.67%6.41%7.69%9.16%N/AN/A6.41%7.69%0.00%N/A26.63%4.44%16.67%7.01%8.41%11.20%N/AN/A7.01%8.41%0.00%N/A100.00%0.00%0.00%cần sửainfeasible –
– e2_videberta_cross_ce_real_frozenemb_lr1e5ViDeBERTa · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 Fsoft-AIC/videberta-baseargmax v1encoderzephyrus 1 1 86M 25.7 ms 23.81%3.97%16.67%6.41%7.69%9.16%N/AN/A6.41%7.69%0.00%N/A26.63%4.44%16.67%7.01%8.41%11.20%N/AN/A7.01%8.41%0.00%N/A100.00%0.00%0.00%cần sửainfeasible –
– e2_videberta_cross_ce_real_frozenemb_lr2e5ViDeBERTa · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 Fsoft-AIC/videberta-baseargmax v1encoderzephyrus 1 1 86M 26.8 ms 23.81%3.97%16.67%6.41%7.69%9.16%N/AN/A6.41%7.69%0.00%N/A26.63%4.44%16.67%7.01%8.41%11.20%N/AN/A7.01%8.41%0.00%N/A100.00%0.00%0.00%cần sửainfeasible –
– e2_videberta_cross_ce_real_frozenemb_lr2e5_fp32 –– v1encoder 0 1 lỗi 0 N/A N/A N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/Afailed –
– e2_videberta_cross_ce_real_frozenemb_lr4e5ViDeBERTa · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 Fsoft-AIC/videberta-baseargmax v1encoderzephyrus 1 1 86M 31.6 ms 23.81%3.97%16.67%6.41%7.69%9.16%N/AN/A6.41%7.69%0.00%N/A26.63%4.44%16.67%7.01%8.41%11.20%N/AN/A7.01%8.41%0.00%N/A100.00%0.00%0.00%cần sửainfeasible –
1 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_lv21mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v2_1 jhu-clsp/mmBERT-base0.076 v2_1encoderzephyrus 3 3 111M 42.4 ms 90.54% ±0.7281.45% ±2.4882.75% ±5.2381.21% ±3.2485.29% ±1.0390.91% ±0.5978.57% ±2.0454.63% ±4.6280.44% ±1.5986.61% ±0.0467.01% ±3.5874.11% ±1.5489.28% ±0.6679.47% ±3.1881.87% ±6.4679.54% ±4.0485.29% ±0.8989.74% ±0.5681.99% ±1.7653.83% ±3.8379.03% ±2.4286.76% ±0.0669.73% ±4.0473.44% ±1.626.46%13.61%4.25%okp85_not_met –
2 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_lv21PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v2_1 vinai/phobert-base-v20.033 v2_1encoderzephyrus 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.

STTModel_NameBackboneThresholdData clean Affirm_Preclean Negate_Preclean Fallback_Preclean Callback_Preclean Relation_Preclean Repeat_Prereply-unseen Affirm_Prereply-unseen Negate_Prereply-unseen Fallback_Prereply-unseen Callback_Prereply-unseen Relation_Prereply-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_m2 –0.051m2 99.38%96.10%56.36%100.00%95.06%81.82%99.45%95.87%55.71%100.00%95.00%60.00%
2 e3_ensemble_devtop1_familygeomean_tempdev_m2 –0.051m2 99.38%96.11%58.22%100.00%95.06%81.82%99.45%95.89%57.35%100.00%95.00%60.00%
3 e3_ensemble_devtop1_familygeomean_temphard_m2 –0.051m2 99.38%96.11%58.22%100.00%95.06%81.82%99.45%95.89%57.35%100.00%95.00%60.00%
4 e3_ensemble_devtop1_familymean_m2 –0.112m2 99.17%95.90%59.42%100.00%95.12%83.33%99.18%95.65%58.59%100.00%95.06%66.67%
5 e3_ensemble_sweep_k2_m2 –0.110m2 98.64%95.61%71.01%96.97%96.20%83.33%98.57%95.34%70.19%96.94%96.15%66.67%
6 e3_ensemble_sweep_k3_m2 –0.051m2 98.97%96.37%63.02%97.00%96.20%69.23%99.01%96.18%61.96%96.97%96.15%42.86%
7 e3_ensemble_sweep_k4_m2 –0.058m2 99.11%96.36%63.02%97.98%96.25%69.23%99.19%96.17%61.96%97.96%96.20%42.86%
8 e3_ensemble_sweep_k5_m2 –0.111m2 99.05%96.24%71.35%97.00%96.25%90.91%99.11%96.05%70.55%96.97%96.20%80.00%
9 e3_ensemble_sweep_k6_m2 –0.087m2 99.11%96.41%67.78%97.00%96.25%90.91%99.10%96.24%67.25%96.97%96.20%80.00%
10 e3_ensemble_sweep_k7_m2 –0.087m2 99.11%96.41%67.78%97.00%96.25%90.91%99.10%96.24%67.25%96.97%96.20%80.00%
11 e3_ensemble_sweep_k8_m2 –0.087m2 99.11%96.41%67.78%97.00%96.25%90.91%99.10%96.24%67.25%96.97%96.20%80.00%
12 e3_ensemble_top_geomean_m2 –0.073m2 99.31%95.75%62.56%97.98%96.34%71.43%99.45%95.50%61.50%97.96%96.30%50.00%
13 e3_ensemble_top_mean_m2 –0.073m2 99.44%95.84%53.02%100.00%96.34%71.43%99.54%95.60%52.49%100.00%96.30%50.00%
14 e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m2 unsloth/gemma-3-1b-pt0.064m2 98.77% ±0.3694.86% ±1.1553.41% ±8.3096.97% ±1.6695.33% ±0.5885.83% ±5.2098.65% ±0.4894.50% ±1.2952.82% ±8.1396.94% ±1.6895.27% ±0.5958.33% ±14.43
15 e2_me5_base_cross_ce_real_frozenemb_lr4e5_m2 intfloat/multilingual-e5-base0.017m2 98.79%94.47%49.38%95.10%92.21%56.25%98.58%94.05%48.07%95.05%92.11%30.00%
16 e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m2 jhu-clsp/mmBERT-base0.155m2 97.71%95.05%63.58%94.90%94.74%71.43%97.35%94.48%62.42%94.85%94.67%50.00%
17 e2_qwen3_17b_base_genlabel_ce_real_lora_m2 Qwen/Qwen3-1.7B-Base0.081m2 99.23%95.11%55.45%96.00%93.83%68.75%99.17%94.77%54.23%95.96%93.75%50.00%
18 e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m2 Qwen/Qwen3-1.7B-Base0.105m2 99.16% ±0.1396.05% ±0.2854.82% ±0.0396.01% ±1.9395.16% ±1.1082.60% ±3.4999.02% ±0.2095.82% ±0.3053.87% ±0.2695.96% ±1.9595.08% ±1.1168.45% ±5.15
19 e3_ensemble_top_greedy_m2 –0.113m2 99.37%95.73%57.89%98.98%96.30%73.33%99.45%95.47%57.00%98.97%96.25%55.56%
20 e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m2 intfloat/multilingual-e5-large-instruct0.017m2 98.66% ±0.2095.35% ±0.2959.78% ±8.7097.95% ±0.0294.41% ±1.5270.89% ±10.4598.51% ±0.2695.05% ±0.3258.92% ±8.5997.92% ±0.0394.34% ±1.5450.22% ±12.00
21 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m2 jhu-clsp/mmBERT-base0.094m2 98.77% ±0.3294.81% ±0.3354.90% ±9.3796.98% ±1.6793.99% ±1.4674.61% ±2.7698.68% ±0.4194.23% ±0.3754.16% ±9.7896.94% ±1.6893.92% ±1.4842.96% ±11.40
22 e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m2 Qwen/Qwen3-4B-Base0.056m2 99.04% ±0.1895.97% ±0.8157.52% ±1.6597.96% ±0.9994.85% ±1.8169.11% ±2.3898.91% ±0.2395.74% ±0.9056.86% ±1.5497.94% ±1.0094.74% ±1.8745.77% ±3.75
23 e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m2 unsloth/gemma-3-1b-pt0.045m2 99.02% ±0.1994.92% ±0.1753.23% ±1.8197.94% ±1.4194.39% ±0.4979.62% ±14.6998.99% ±0.2694.56% ±0.1952.45% ±1.8797.92% ±1.4394.32% ±0.4958.93% ±22.73
24 e2_gemma3_1b_pt_cross_ce_real_lora_r32_m2 unsloth/gemma-3-1b-pt0.045m2 98.68% ±0.1994.78% ±1.2351.30% ±6.0598.27% ±1.1592.64% ±2.5471.69% ±10.6398.59% ±0.2194.32% ±1.2350.48% ±6.2498.25% ±1.1692.93% ±2.7049.84% ±14.64
25 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m2 jhu-clsp/mmBERT-base0.052m2 98.47% ±0.1494.82% ±0.0754.21% ±4.2395.23% ±0.4593.24% ±1.3161.19% ±8.9398.26% ±0.1894.37% ±0.1254.50% ±3.5095.18% ±0.4593.15% ±1.3242.86% ±7.14
26 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m2 vinai/phobert-base-v20.008m2 98.76%94.85%56.13%95.96%91.46%100.00%98.73%94.72%57.36%95.92%91.36%100.00%
27 e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m2 Qwen/Qwen3-4B-Base0.162m2 99.03%95.68%54.95%97.00%92.94%70.59%99.00%95.40%54.03%96.97%92.86%54.55%
28 e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m2 Qwen/Qwen3-8B-Base0.170m2 98.84%96.34%66.29%96.00%90.80%73.33%98.74%96.16%65.48%95.96%90.70%55.56%
29 e3_ensemble_seed6_kd_a05_t1_m2 –0.067m2 98.76%95.68%60.80%96.04%95.12%90.00%98.64%95.42%60.32%96.00%95.06%75.00%
30 e2_me5_base_cross_ce_real_frozenemb_lr2e5_m2 intfloat/multilingual-e5-base0.031m2 98.88% ±0.0694.36% ±0.5555.47% ±5.8294.76% ±0.5592.56% ±1.2571.82% ±1.5798.73% ±0.1393.87% ±0.6454.39% ±5.9194.70% ±0.5692.47% ±1.2735.00% ±8.66
31 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m2 jhu-clsp/mmBERT-base0.061m2 98.40% ±0.4195.31% ±0.8953.62% ±6.4598.00% ±2.6192.80% ±1.0368.81% ±7.8498.23% ±0.4994.91% ±0.8453.19% ±6.0297.98% ±2.6392.70% ±1.0444.44% ±9.62
32 e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m2 Qwen/Qwen3-4B-Base0.131m2 99.07% ±0.1395.54% ±0.4763.85% ±10.4097.38% ±2.9695.19% ±1.9574.85% ±22.2198.98% ±0.1295.26% ±0.5263.18% ±10.6597.35% ±2.9895.13% ±1.9760.97% ±33.93
33 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.043m2 98.64% ±0.2294.37% ±0.5754.24% ±2.2296.98% ±0.9794.25% ±1.3083.33% ±16.6798.47% ±0.3293.90% ±0.5453.56% ±2.0696.95% ±0.9894.18% ±1.3266.67% ±33.33
34 e2_gemma3_1b_pt_cross_ce_real_lora_m2 unsloth/gemma-3-1b-pt0.053m2 98.64% ±0.2894.73% ±0.7155.38% ±3.1197.26% ±1.5294.94% ±1.0571.78% ±9.0498.51% ±0.4294.28% ±0.6955.26% ±2.8697.23% ±1.5494.88% ±1.0646.79% ±11.75
35 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m2 vinai/phobert-base-v20.085m2 98.34% ±0.4095.41% ±0.4361.33% ±2.2795.73% ±0.5492.14% ±4.0478.21% ±5.3198.18% ±0.5394.96% ±0.3560.99% ±2.7895.68% ±0.5592.38% ±3.5156.39% ±14.35
36 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m2 jhu-clsp/mmBERT-base0.082m2 98.50% ±0.2394.81% ±0.7156.84% ±8.2795.60% ±0.5291.96% ±1.8163.10% ±1.0398.27% ±0.2594.28% ±0.6556.07% ±8.3495.55% ±0.5291.85% ±1.8439.17% ±1.44
37 e2_qwen3_17b_base_cross_ce_real_lora_m2 Qwen/Qwen3-1.7B-Base0.101m2 98.63% ±0.2094.81% ±0.7354.65% ±5.0297.30% ±2.0693.63% ±0.5364.35% ±4.3898.47% ±0.2694.44% ±0.8154.15% ±5.0097.27% ±2.0893.56% ±0.5340.28% ±8.67
38 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m2 vinai/phobert-base-v20.068m2 98.74% ±0.3694.94% ±0.4058.47% ±5.0396.20% ±0.9393.79% ±1.2884.54% ±9.1098.61% ±0.4794.53% ±0.4458.21% ±5.1696.16% ±0.9493.91% ±1.1767.54% ±17.56
39 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m2 vinai/phobert-base-v20.052m2 98.82% ±0.3594.72% ±2.3955.84% ±4.3597.63% ±0.5593.58% ±2.4078.32% ±9.1698.69% ±0.5194.20% ±2.5155.31% ±4.7697.61% ±0.5593.82% ±1.8754.07% ±13.39
40 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m2 vinai/phobert-base-v20.058m2 98.68% ±0.3693.92% ±0.7756.84% ±2.8096.96% ±0.9591.98% ±2.2473.59% ±9.3498.57% ±0.4593.53% ±0.7456.28% ±3.2996.93% ±0.9691.88% ±2.2650.71% ±14.09
41 e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m2 unsloth/gemma-3-1b-pt0.056m2 98.66% ±0.1495.15% ±1.1153.03% ±3.7196.61% ±1.4895.67% ±0.8870.69% ±11.5298.51% ±0.1394.97% ±1.2452.42% ±3.0996.57% ±1.5095.61% ±0.8948.79% ±11.86
42 e2_gemma3_270m_cross_ce_real_lora_m2 unsloth/gemma-3-270m0.098m2 97.95% ±0.6194.12% ±0.5151.17% ±2.9694.31% ±1.0794.21% ±1.8557.94% ±2.6197.57% ±0.8093.64% ±0.7650.72% ±3.3794.25% ±1.0894.13% ±1.8735.14% ±1.60
43 e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m2 intfloat/multilingual-e5-large-instruct0.029m2 98.69% ±0.3095.45% ±0.6657.94% ±1.3097.30% ±1.5394.03% ±1.4176.09% ±11.4998.64% ±0.3995.24% ±0.6057.30% ±1.4397.27% ±1.5593.95% ±1.4253.70% ±11.56
44 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m2 jhu-clsp/mmBERT-base0.040m2 98.60% ±0.4095.60% ±0.4550.75% ±4.6196.27% ±1.0692.61% ±2.9564.97% ±10.5698.39% ±0.4595.16% ±0.4849.87% ±4.2596.23% ±1.0692.91% ±2.3236.71% ±8.30
45 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m2 jhu-clsp/mmBERT-base0.057m2 98.77% ±0.2794.96% ±0.4154.48% ±4.9998.60% ±0.6193.68% ±1.2162.70% ±11.2598.62% ±0.3694.45% ±0.3754.35% ±3.5998.59% ±0.6293.60% ±1.2242.86% ±12.37
46 e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m2 jhu-clsp/mmBERT-base0.070m2 98.28%95.52%59.90%95.96%93.51%73.33%98.10%95.01%60.00%95.92%93.42%55.56%
47 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m2 jhu-clsp/mmBERT-base0.047m2 98.38% ±0.2094.91% ±0.5054.56% ±2.6495.62% ±1.4192.89% ±0.5668.86% ±14.4298.17% ±0.2394.48% ±0.5954.08% ±2.8895.57% ±1.4292.80% ±0.5744.07% ±19.57
48 e2_xlmr_base_khonly_ce_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.012m2 98.25%94.89%52.40%91.51%91.57%72.73%98.15%94.31%51.60%91.43%91.46%40.00%
49 e3_ensemble_seed_kd_a05_t1_m2 –0.078m2 98.56%95.46%60.50%97.00%95.18%90.00%98.46%95.19%60.32%96.97%95.12%75.00%
50 e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m2 unsloth/gemma-3-1b-pt0.042m2 98.42% ±0.1994.97% ±0.5751.46% ±2.4097.31% ±2.5293.93% ±1.9283.46% ±5.8998.20% ±0.2594.61% ±0.6451.37% ±2.0397.29% ±2.5593.85% ±1.9565.83% ±8.04
51 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m2 vinai/phobert-base-v20.024m2 98.75% ±0.4595.05% ±0.1851.89% ±0.9396.69% ±1.1294.21% ±1.2286.60% ±5.8098.62% ±0.5094.65% ±0.2851.31% ±0.8396.65% ±1.1394.13% ±1.2365.56% ±15.03
52 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m2 vinai/phobert-base-v20.031m2 98.26% ±0.3195.88% ±0.1654.30% ±5.6497.31% ±1.1093.38% ±2.4577.78% ±3.8598.07% ±0.4195.64% ±0.1853.41% ±5.4997.28% ±1.1193.30% ±2.4751.85% ±3.21
53 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t2_tsa_m2 vinai/phobert-base-v20.070m2 98.39% ±0.1094.92% ±0.2457.43% ±3.3596.33% ±0.5893.88% ±1.1982.95% ±8.9198.24% ±0.1494.36% ±0.2557.06% ±3.7496.30% ±0.5893.77% ±1.2468.99% ±12.40
54 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m2 vinai/phobert-base-v20.031m2 98.66% ±0.2794.80% ±0.4456.35% ±4.9396.00% ±0.9294.05% ±0.8884.44% ±5.0998.57% ±0.3794.37% ±0.6055.69% ±5.1195.96% ±0.9393.97% ±0.9063.89% ±12.73
55 e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m2 unsloth/gemma-3-1b-pt0.096m2 98.62% ±0.1093.48% ±0.3755.37% ±6.6297.62% ±2.1294.22% ±1.4150.32% ±15.7098.42% ±0.0893.05% ±0.5055.75% ±7.3097.59% ±2.1594.14% ±1.4328.24% ±8.14
56 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m2 intfloat/multilingual-e5-large-instruct0.038m2 98.73% ±0.3795.08% ±0.3552.92% ±3.9897.96% ±0.9792.25% ±1.7572.62% ±2.0698.62% ±0.4594.82% ±0.3452.25% ±3.7797.94% ±0.9892.16% ±1.7750.00% ±0.00
57 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.017m2 98.37% ±0.0495.38% ±0.0956.83% ±1.8098.29% ±0.5992.69% ±1.9688.43% ±11.1498.16% ±0.0595.08% ±0.2056.50% ±2.0398.27% ±0.6092.60% ±1.9861.11% ±34.69
58 e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m2 vinai/phobert-base-v20.026m2 98.38% ±0.2194.62% ±0.7361.41% ±6.2196.99% ±1.7292.98% ±1.4080.56% ±7.3598.23% ±0.2894.39% ±0.8760.97% ±5.8896.96% ±1.7492.89% ±1.4250.00% ±16.67
59 e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.065m2 98.57% ±0.3694.69% ±0.1855.44% ±2.8595.67% ±1.4694.19% ±0.7074.16% ±6.7298.42% ±0.5094.24% ±0.2655.05% ±3.4595.62% ±1.4794.12% ±0.7150.95% ±8.61
60 e2_xlmr_base_cross_logitadjust_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.018m2 98.44% ±0.0795.37% ±0.2949.47% ±2.5894.16% ±1.8590.27% ±3.2067.58% ±2.8698.15% ±0.1794.99% ±0.2748.88% ±3.0194.10% ±1.8690.43% ±2.7641.07% ±3.09
61 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m2 intfloat/multilingual-e5-large-instruct0.026m2 98.75% ±0.2095.10% ±0.3962.23% ±6.6296.95% ±0.9495.40% ±1.3066.67% ±0.0098.57% ±0.2294.77% ±0.4361.31% ±6.5296.92% ±0.9595.34% ±1.3244.44% ±0.00
62 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m2 jhu-clsp/mmBERT-base0.072m2 98.10% ±0.3194.89% ±0.6458.33% ±3.9395.94% ±1.5892.64% ±1.0364.66% ±9.8797.83% ±0.3794.38% ±0.8157.47% ±3.6995.90% ±1.6092.54% ±1.0538.36% ±14.42
63 e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m2 jhu-clsp/mmBERT-base0.082m2 98.25% ±0.6295.21% ±0.5456.47% ±1.2797.20% ±0.6193.17% ±1.4160.83% ±6.8297.97% ±0.7494.72% ±0.5155.89% ±1.4197.17% ±0.6193.08% ±1.4235.56% ±11.76
64 e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m2 Qwen/Qwen3-1.7B-Base0.067m2 98.66%94.90%47.18%94.12%96.15%76.92%98.51%94.53%46.61%94.06%96.10%57.14%
65 e2_visobert_cross_ce_real_frozenemb_lr4e5_m2 uitnlp/visobert0.037m2 98.70%93.57%40.83%94.85%93.59%69.23%98.56%93.27%40.96%94.79%93.51%50.00%
66 e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m2 FacebookAI/xlm-roberta-base0.027m2 98.02%94.46%49.56%95.10%92.59%69.23%97.76%94.05%49.54%95.05%92.50%42.86%
67 e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m2 jhu-clsp/mmBERT-base0.057m2 98.40% ±0.0794.56% ±0.7553.35% ±5.2098.61% ±0.5898.26% ±1.5190.24% ±9.1798.23% ±0.1194.17% ±0.8353.24% ±4.1698.60% ±0.5998.23% ±1.5375.56% ±21.43
68 e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.034m2 98.47% ±0.1094.62% ±0.6060.55% ±4.0096.05% ±0.9591.73% ±1.8882.02% ±15.5798.25% ±0.0594.02% ±0.6659.96% ±4.3996.01% ±0.9691.57% ±1.8965.19% ±31.14
69 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m2 jhu-clsp/mmBERT-base0.144m2 98.35% ±0.2594.86% ±0.2554.35% ±3.5897.18% ±0.5492.18% ±0.6464.48% ±2.0998.16% ±0.3894.35% ±0.1853.16% ±3.5197.15% ±0.5492.47% ±0.1536.94% ±3.37
70 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m2 jhu-clsp/mmBERT-base0.108m2 98.27% ±0.6195.09% ±0.4661.36% ±5.7597.29% ±2.3492.37% ±0.8967.90% ±7.8698.06% ±0.7694.68% ±0.5360.52% ±5.6597.26% ±2.3792.27% ±0.9044.88% ±10.69
71 e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m2 Qwen/Qwen3-0.6B-Base0.075m2 98.68% ±0.2395.57% ±0.4153.36% ±6.7395.40% ±1.4595.49% ±1.4464.62% ±2.0898.47% ±0.2595.19% ±0.4452.78% ±6.9695.35% ±1.4695.43% ±1.4644.78% ±0.58
72 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t1_tsa_m2 vinai/phobert-base-v20.070m2 98.43% ±0.2294.98% ±0.5256.35% ±3.2496.04% ±0.9493.77% ±2.0182.27% ±9.1898.21% ±0.3294.42% ±0.5655.81% ±3.5396.00% ±0.9493.70% ±2.0366.30% ±14.92
73 e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m2 Qwen/Qwen3-0.6B-Base0.077m2 97.84% ±0.2093.54% ±0.6147.85% ±5.3198.26% ±1.2096.61% ±0.8080.19% ±9.5297.45% ±0.2593.02% ±0.6750.09% ±8.0198.25% ±1.2296.57% ±0.8159.05% ±20.07
74 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m2 vinai/phobert-base-v20.104m2 98.36% ±0.5293.98% ±0.5061.64% ±9.1396.35% ±1.1394.35% ±0.9682.78% ±15.1698.05% ±0.6293.39% ±0.6861.18% ±9.2696.31% ±1.1494.27% ±0.9869.05% ±27.04
75 e2_xlmr_base_cross_ce_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.044m2 98.15% ±0.3994.51% ±0.5353.13% ±8.8394.71% ±0.5190.53% ±1.0269.80% ±3.0497.89% ±0.5494.19% ±0.4452.71% ±9.3794.65% ±0.5290.41% ±1.0336.48% ±10.19
76 e2_xlmr_base_cross_wcesqrt_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.019m2 98.18% ±0.5794.61% ±0.6654.14% ±6.0994.76% ±1.0790.89% ±2.3568.35% ±5.4797.90% ±0.7094.13% ±0.7053.26% ±6.1794.71% ±1.0890.78% ±2.3746.14% ±8.28
77 e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.049m2 98.57% ±0.6395.09% ±0.5354.85% ±0.9596.34% ±1.4492.63% ±2.1266.87% ±4.4798.42% ±0.7894.68% ±0.4654.50% ±1.3796.30% ±1.4592.53% ±2.1441.11% ±8.39
78 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m2 vinai/phobert-base-v20.159m2 97.99% ±0.3094.23% ±0.8059.12% ±8.4596.35% ±1.1393.67% ±0.9670.64% ±9.2797.70% ±0.4194.02% ±0.5558.97% ±7.7896.31% ±1.1493.59% ±0.9752.38% ±4.12
79 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsa_m2 vinai/phobert-base-v20.085m2 98.35% ±0.3195.35% ±0.1559.44% ±5.3496.04% ±0.0093.10% ±1.2672.80% ±4.0998.10% ±0.3194.83% ±0.1758.99% ±5.3996.00% ±0.0093.02% ±1.2849.05% ±8.61
80 e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m2 Qwen/Qwen3-Embedding-4B0.061m2 96.87% ±0.6092.84% ±0.1150.76% ±0.7698.26% ±0.9590.44% ±5.4467.92% ±31.2396.07% ±0.9092.11% ±0.0349.53% ±1.3098.24% ±0.9690.23% ±5.6251.39% ±33.39
81 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m2 vinai/phobert-base-v20.025m2 97.96%94.41%58.85%96.00%91.03%50.00%97.68%94.02%60.11%95.96%90.91%33.33%
82 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m2 Qwen/Qwen3-Embedding-0.6B0.152m2 97.21% ±0.6291.73% ±1.3555.34% ±5.4693.26% ±1.4590.05% ±1.5158.75% ±5.3496.54% ±0.8291.00% ±1.4954.02% ±5.8093.19% ±1.4689.91% ±1.5237.48% ±5.87
83 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m2 vinai/phobert-base-v20.035m2 98.68% ±0.1295.23% ±0.9856.20% ±6.1197.31% ±1.1392.83% ±3.6476.08% ±5.2898.60% ±0.2194.83% ±0.9655.26% ±6.0397.28% ±1.1492.75% ±3.6853.33% ±5.77
84 e2_me5_base_cross_ce_real_frozenemb_lr1e5_m2 intfloat/multilingual-e5-base0.052m2 98.18%94.21%51.11%92.23%94.87%50.00%97.88%93.76%51.67%92.16%94.81%40.00%
85 e2_visobert_cross_ce_real_frozenemb_lr2e5_m2 uitnlp/visobert0.027m2 97.74%92.53%43.51%95.88%93.75%58.33%97.39%92.33%43.37%95.83%93.67%16.67%
86 e2_visobert_cross_ce_real_frozenemb_lr1e5_m2 uitnlp/visobert0.186m2 97.33%90.91%45.61%95.88%89.19%57.14%96.67%90.34%45.33%95.83%90.14%25.00%
87 e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m2 FacebookAI/xlm-roberta-base0.161m2 97.17%94.07%54.55%92.23%91.14%46.67%96.63%93.63%54.00%92.16%91.03%11.11%
88 e1_tfidf_hybrid_qwen3emb_lr_m2 Qwen/Qwen3-Embedding-0.6B0.149m2 96.65%89.80%49.31%93.62%93.15%72.73%95.97%88.61%49.26%93.55%93.06%40.00%
89 e1_tfidf_svmcal_ce_real_base_m2 TF-IDF0.178m2 97.46%89.14%57.92%94.85%94.59%72.73%96.93%87.95%57.56%94.79%94.52%40.00%
90 e2_dual_qwen3emb06b_concat_ce_real_frozen_m2 Qwen/Qwen3-Embedding-0.6B0.227m2 95.19% ±0.8890.03% ±0.9547.17% ±3.5195.42% ±0.4585.32% ±1.9192.96% ±6.1294.03% ±1.0889.24% ±0.9346.39% ±3.4195.37% ±0.4685.11% ±1.9280.56% ±17.35
– e3_fbverifier_devtop1_familygeomean_fbf1_m2 –hardm2 98.59%95.16%87.20%96.04%95.12%64.71%98.51%95.07%87.29%96.00%95.06%45.45%
– e3_fbverifier_devtop1_familygeomean_macrof1_m2 –hardm2 98.59%95.16%87.20%96.04%95.12%64.71%98.51%95.07%87.29%96.00%95.06%45.45%
1 e3_ensemble_devtop1_familygeomean_m1ensemble trung bình nhân theo nhóm trên 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051m1 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 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051m1 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 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051m1 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_m1ensemble trung bình đều theo nhóm trên 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.112m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 2 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.110m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 3 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 4 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.058m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.111m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.087m1 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 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.087m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.087m1 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_m1ensemble trung bình nhân trên 5 phương pháp · nhãn m1 –0.073m1 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_m1ensemble trung bình trên 5 phương pháp · nhãn m1 –0.073m1 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_m1gemma-3-1b-pt · 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 unsloth/gemma-3-1b-pt0.064m1 98.29% ±0.2694.13% ±0.9951.43% ±7.9392.63% ±2.6593.88% ±0.9576.39% ±10.4998.02% ±0.3493.70% ±1.1150.78% ±7.7592.56% ±2.6793.80% ±0.9644.44% ±9.62
15 e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-base0.017m1 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_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.155m1 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_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-1.7B-Base0.081m1 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_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-1.7B-Base0.105m1 98.68% ±0.1995.08% ±0.4251.49% ±0.4392.39% ±2.6493.07% ±0.9682.60% ±3.4998.42% ±0.2494.60% ±0.5150.68% ±0.7892.00% ±2.2192.96% ±0.9768.45% ±5.15
19 e3_ensemble_top_greedy_m1ensemble chọn tham lam Caruana trên 5 phương pháp · nhãn m1 –0.113m1 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_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 3e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.017m1 98.19% ±0.2394.13% ±0.2956.58% ±7.5692.58% ±0.5494.21% ±0.7569.06% ±9.2897.90% ±0.3093.70% ±0.3255.64% ±7.4292.51% ±0.5594.14% ±0.7647.84% ±10.56
21 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m1 jhu-clsp/mmBERT-base0.094m1 98.41% ±0.3993.54% ±0.3751.61% ±8.9194.11% ±1.8393.78% ±1.2474.61% ±2.7698.21% ±0.5292.83% ±0.4150.79% ±9.2494.05% ±1.8493.70% ±1.2542.96% ±11.40
22 e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1factorized · train thật + synth ×1, nhãn m1 jhu-clsp/mmBERT-base0.108m1 98.11% ±0.3693.84% ±1.1746.27% ±6.8090.14% ±0.9890.77% ±2.1573.53% ±2.5597.73% ±0.4493.30% ±1.2346.89% ±6.3890.04% ±0.9890.66% ±2.1651.52% ±2.62
23 e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-4B-Base0.056m1 98.57% ±0.2695.16% ±0.6254.20% ±1.9394.48% ±1.3993.54% ±0.6264.48% ±2.0998.30% ±0.3494.91% ±0.6253.35% ±1.9394.43% ±1.4093.41% ±0.6840.65% ±3.52
24 e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 5 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.045m1 98.58% ±0.2494.06% ±0.0150.21% ±1.1594.51% ±1.9492.90% ±1.0979.62% ±14.6998.40% ±0.3193.50% ±0.1749.49% ±1.6494.45% ±1.9592.82% ±1.0958.93% ±22.73
25 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-v20.033m1 98.27% ±0.3093.63% ±0.3148.47% ±2.3292.12% ±2.5490.65% ±1.2972.88% ±8.3998.06% ±0.3493.07% ±0.3248.25% ±2.5492.04% ±2.5690.90% ±1.5140.00% ±17.32
26 e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=32 · CE · lr 0.0001, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.045m1 98.38% ±0.1993.47% ±1.0048.73% ±5.8295.31% ±2.4190.19% ±2.8770.04% ±11.5598.20% ±0.2193.02% ±1.0047.68% ±6.0595.57% ±2.0490.41% ±2.9548.06% ±16.17
27 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 6 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.052m1 98.05% ±0.1493.59% ±0.3151.25% ±3.9890.93% ±0.7993.05% ±1.5259.53% ±10.4097.71% ±0.1893.00% ±0.4951.33% ±3.3790.84% ±0.7992.97% ±1.5441.07% ±7.78
28 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-base0.051m1 98.20% ±0.1494.27% ±0.2953.27% ±4.3094.33% ±1.4393.17% ±2.1170.48% ±3.4397.93% ±0.1493.78% ±0.3952.44% ±3.9294.27% ±1.4493.08% ±2.1350.00% ±5.56
29 e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 jhu-clsp/mmBERT-base0.118m1 98.29% ±0.5093.98% ±0.0652.13% ±6.7493.70% ±1.4791.69% ±0.8570.09% ±8.9298.09% ±0.6393.30% ±0.0651.28% ±7.0193.63% ±1.4891.59% ±0.8648.68% ±13.31
30 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1PhoBERT-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 m1 vinai/phobert-base-v20.008m1 98.22%93.10%53.07%91.35%89.53%100.00%97.93%92.77%54.25%91.26%89.41%100.00%
31 e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1 Qwen/Qwen3-4B-Base0.162m1 98.70%95.25%51.44%93.33%89.25%66.67%98.56%94.91%50.43%93.27%89.13%50.00%
32 e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1 Qwen/Qwen3-8B-Base0.170m1 98.44%95.36%61.86%93.27%89.25%68.75%98.23%95.07%60.87%93.20%89.13%50.00%
33 e3_ensemble_seed6_kd_a05_t1_m1ensemble trung bình nhân trên 1 phương pháp · nhãn m1 –0.067m1 98.30%94.49%57.48%90.65%94.12%90.00%98.04%94.10%56.86%90.57%94.05%75.00%
34 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_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-v20.073m1 98.07% ±0.1395.16% ±0.4453.03% ±1.2792.67% ±1.2988.54% ±5.4570.46% ±8.3297.80% ±0.1394.59% ±0.5052.74% ±1.0692.60% ±1.3088.41% ±5.5046.63% ±2.96
35 e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 unsloth/gemma-3-1b-pt0.064m1 98.29% ±0.6393.22% ±0.4554.46% ±1.4094.09% ±1.8492.50% ±0.8267.86% ±6.1997.97% ±0.8592.76% ±0.4353.74% ±1.3594.03% ±1.8692.41% ±0.8241.67% ±7.22
36 e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-base0.031m1 98.49% ±0.0793.20% ±0.5352.23% ±5.4189.76% ±0.8491.56% ±0.2769.70% ±5.2598.23% ±0.1292.58% ±0.5951.03% ±5.5389.66% ±0.8591.46% ±0.2833.33% ±11.55
37 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.061m1 97.98% ±0.4393.99% ±1.0950.36% ±5.3894.75% ±2.9292.22% ±0.2768.81% ±7.8497.68% ±0.5293.45% ±1.0649.81% ±5.0094.70% ±2.9592.12% ±0.2744.44% ±9.62
38 e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1 Qwen/Qwen3-4B-Base0.131m1 98.53% ±0.1394.56% ±0.4460.20% ±9.1693.09% ±3.5893.47% ±2.2369.47% ±18.9498.28% ±0.1194.17% ±0.4859.47% ±9.2093.02% ±3.6193.39% ±2.2551.90% ±24.43
39 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_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 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.043m1 98.10% ±0.2193.07% ±0.4451.47% ±2.3992.94% ±1.8392.55% ±1.5377.78% ±19.2597.77% ±0.3192.48% ±0.3850.77% ±2.3392.87% ±1.8492.46% ±1.5559.26% ±35.72
40 e2_gemma3_1b_pt_cross_ce_real_lora_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.053m1 98.17% ±0.2893.59% ±0.5152.69% ±2.8794.34% ±1.4092.36% ±1.2271.78% ±9.0497.90% ±0.4293.04% ±0.4852.60% ±2.6494.28% ±1.4192.26% ±1.2346.79% ±11.75
41 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_kd_a05_t1_tsa_m1 vinai/phobert-base-v20.056m1 98.25% ±0.3493.65% ±0.6851.62% ±1.0292.61% ±1.0292.11% ±2.5271.02% ±6.1297.97% ±0.3693.09% ±0.6450.96% ±1.1592.54% ±1.0392.02% ±2.5550.87% ±5.89
42 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-v20.037m1 98.11% ±0.0893.21% ±1.4153.11% ±2.2291.41% ±2.3192.94% ±0.1483.42% ±14.4797.78% ±0.0292.67% ±1.5652.59% ±2.4191.33% ±2.3392.86% ±0.1570.90% ±25.21
43 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_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, nhãn m1 vinai/phobert-base-v20.085m1 97.82% ±0.4694.22% ±0.4258.20% ±2.0890.37% ±0.4890.63% ±4.1074.94% ±7.5497.49% ±0.6193.63% ±0.3457.72% ±2.5490.28% ±0.4990.83% ±3.6252.22% ±13.47
44 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_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, nhãn m1 vinai/phobert-base-v20.028m1 98.31% ±0.2093.54% ±1.3350.22% ±1.2694.12% ±1.5792.78% ±1.1760.00% ±9.2198.10% ±0.3093.05% ±1.4749.43% ±1.0194.06% ±1.5892.69% ±1.1834.05% ±20.14
45 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.082m1 98.06% ±0.2293.31% ±0.8053.54% ±7.3792.78% ±1.3290.68% ±2.2959.55% ±4.4197.69% ±0.2592.63% ±0.7852.68% ±7.3992.71% ±1.3390.56% ±2.3135.73% ±2.15
46 e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1, nhãn m1 jhu-clsp/mmBERT-base0.057m1 98.34% ±0.0694.20% ±0.6148.07% ±1.1094.03% ±1.7492.36% ±2.5662.08% ±6.8898.11% ±0.1393.62% ±0.7347.62% ±1.4793.97% ±1.7692.26% ±2.5841.90% ±7.33
47 e2_qwen3_17b_base_cross_ce_real_lora_m1Qwen3-1.7B Base · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật, nhãn m1 Qwen/Qwen3-1.7B-Base0.101m1 98.03% ±0.3594.04% ±0.7151.55% ±4.4693.53% ±2.3692.67% ±0.4764.35% ±4.3897.68% ±0.4693.58% ±0.7950.94% ±4.4193.46% ±2.3892.59% ±0.4840.28% ±8.67
48 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1 vinai/phobert-base-v20.068m1 98.28% ±0.3793.86% ±0.5755.47% ±4.9791.09% ±1.1792.72% ±1.5184.54% ±9.1098.02% ±0.4993.33% ±0.6355.07% ±5.0891.01% ±1.1892.82% ±1.4567.54% ±17.56
49 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_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, nhãn m1 vinai/phobert-base-v20.052m1 98.33% ±0.3393.27% ±2.3552.32% ±4.1993.21% ±0.8892.32% ±2.8776.79% ±10.6798.04% ±0.4892.60% ±2.4551.65% ±4.5293.15% ±0.8892.53% ±2.3852.22% ±13.47
50 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-v20.058m1 98.17% ±0.3492.37% ±0.7253.92% ±2.7093.20% ±1.7490.45% ±2.4562.66% ±8.2797.89% ±0.4291.82% ±0.6953.26% ±3.1493.13% ±1.7590.34% ±2.4738.10% ±10.91
51 e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.056m1 98.28% ±0.1294.03% ±1.2450.62% ±3.7491.63% ±1.2993.45% ±0.9270.69% ±11.5298.02% ±0.0993.74% ±1.4050.05% ±3.4491.55% ±1.3093.37% ±0.9348.79% ±11.86
52 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-3-270m0.098m1 97.42% ±0.7493.08% ±0.9148.54% ±2.5390.94% ±0.4893.14% ±1.7855.98% ±3.8396.86% ±1.0092.51% ±1.2248.16% ±3.1891.45% ±0.4893.05% ±1.8033.46% ±4.95
53 e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.029m1 98.27% ±0.2294.16% ±0.9054.45% ±1.2993.56% ±1.4293.81% ±1.1571.74% ±8.7998.08% ±0.2993.80% ±0.8753.81% ±1.5693.50% ±1.4393.74% ±1.1646.67% ±5.77
54 e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.034m1 98.50% ±0.2293.44% ±0.4747.86% ±2.3792.37% ±3.5292.70% ±1.8368.74% ±9.4398.28% ±0.3192.93% ±0.5249.72% ±0.5392.30% ±3.5592.61% ±1.8551.34% ±11.37
55 e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.051m1 98.10% ±0.0493.55% ±1.2249.95% ±0.9794.89% ±1.4791.38% ±1.3360.73% ±15.5197.81% ±0.0993.12% ±1.3750.45% ±0.7694.84% ±1.4991.27% ±1.3542.54% ±17.01
56 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.040m1 98.24% ±0.4094.20% ±0.4647.82% ±4.1593.74% ±2.0092.03% ±3.1062.95% ±8.7697.92% ±0.4693.61% ±0.4446.88% ±3.8393.68% ±2.0292.32% ±2.4834.49% ±7.86
57 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.057m1 98.27% ±0.2693.39% ±0.5551.27% ±4.5294.98% ±1.0393.02% ±0.8362.70% ±11.2597.98% ±0.3592.72% ±0.5550.96% ±3.3094.93% ±1.0492.94% ±0.8442.86% ±12.37
58 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-base0.070m1 97.95%94.35%56.60%90.48%92.59%64.71%97.67%93.72%56.50%90.38%92.50%45.45%
59 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1factorized · train thật, nhãn m1 jhu-clsp/mmBERT-base0.047m1 97.98% ±0.2393.61% ±0.4050.94% ±2.2692.84% ±2.1593.04% ±0.5968.86% ±14.4297.65% ±0.2693.04% ±0.5050.31% ±2.2892.77% ±2.1792.95% ±0.6044.07% ±19.57
60 e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1gen_label · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 Qwen/Qwen3-1.7B-Base0.095m1 98.41% ±0.3894.09% ±0.6054.11% ±4.0196.29% ±0.5493.79% ±2.2183.60% ±5.1698.06% ±0.4893.56% ±0.5552.95% ±4.1796.25% ±0.5493.72% ±2.2368.06% ±6.36
61 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-base0.012m1 97.72%93.90%50.62%86.61%91.86%66.67%97.44%93.22%49.57%86.49%91.76%28.57%
62 e3_ensemble_seed_kd_a05_t1_m1ensemble trung bình nhân trên 1 phương pháp · nhãn m1 –0.078m1 98.16%94.30%57.75%92.38%94.25%81.82%97.94%93.91%57.43%92.31%94.19%60.00%
63 e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1gemma-3-1b-pt · 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 unsloth/gemma-3-1b-pt0.042m1 97.93% ±0.1694.42% ±0.7049.59% ±2.8492.84% ±2.5991.75% ±1.7778.46% ±3.4197.56% ±0.2194.01% ±0.8049.53% ±2.4192.77% ±2.6191.65% ±1.8057.50% ±6.61
64 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_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, nhãn m1 vinai/phobert-base-v20.024m1 98.28% ±0.4793.71% ±0.1649.35% ±1.0092.45% ±1.4992.80% ±1.2681.94% ±6.2198.00% ±0.5393.17% ±0.1648.60% ±0.8192.38% ±1.5192.72% ±1.2857.94% ±8.36
65 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_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 2e-05, 6 epoch · train thật, nhãn m1 vinai/phobert-base-v20.031m1 97.74% ±0.3294.57% ±0.1951.77% ±5.1292.64% ±1.7692.11% ±2.8277.78% ±3.8597.41% ±0.4194.19% ±0.2350.93% ±5.1092.57% ±1.7792.02% ±2.8551.85% ±3.21
66 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t2_tsa_m1 vinai/phobert-base-v20.070m1 97.93% ±0.1093.89% ±0.1554.68% ±3.3991.75% ±0.5593.31% ±1.9178.90% ±8.8197.64% ±0.1393.22% ±0.1854.21% ±3.7391.67% ±0.5693.20% ±1.9462.70% ±11.25
67 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_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-v20.031m1 98.15% ±0.2993.61% ±0.3853.44% ±4.4791.13% ±1.2292.31% ±1.7979.58% ±2.4797.90% ±0.3893.06% ±0.5052.69% ±4.6191.05% ±1.2392.22% ±1.8155.71% ±5.15
68 e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 5e-05, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.096m1 98.17% ±0.1492.59% ±0.5053.06% ±6.0694.08% ±2.4892.19% ±1.3145.76% ±16.7797.78% ±0.1792.08% ±0.5753.09% ±6.5794.02% ±2.5092.06% ±1.3924.44% ±7.70
69 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.038m1 98.16% ±0.4094.08% ±0.6049.81% ±3.2593.52% ±0.9391.23% ±0.7471.03% ±4.1897.88% ±0.4893.70% ±0.4649.08% ±3.0493.46% ±0.9391.12% ±0.7548.15% ±3.21
70 e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1factorized · train thật + synth ×1, nhãn m1 jhu-clsp/mmBERT-base0.061m1 98.16% ±0.6193.85% ±0.2747.34% ±1.9792.93% ±3.4192.34% ±1.6468.50% ±15.6397.92% ±0.7693.24% ±0.1947.88% ±2.7292.86% ±3.4492.65% ±2.1943.89% ±21.10
71 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.017m1 97.95% ±0.0693.87% ±0.1253.52% ±1.5894.11% ±0.9091.07% ±2.4788.43% ±11.1497.58% ±0.0893.40% ±0.2953.00% ±1.8494.05% ±0.9190.96% ±2.4961.11% ±34.69
72 e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.026m1 97.99% ±0.2693.08% ±0.7757.55% ±6.0592.64% ±1.4091.01% ±2.2380.56% ±7.3597.71% ±0.3492.68% ±0.9557.07% ±5.6992.57% ±1.4290.91% ±2.2550.00% ±16.67
73 e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_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 MLP 512 · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.065m1 98.02% ±0.4093.45% ±0.3752.98% ±2.8191.40% ±1.7392.47% ±1.3169.49% ±11.1897.70% ±0.5492.87% ±0.4752.41% ±3.4791.32% ±1.7492.34% ±1.3545.93% ±13.39
74 e2_xlmr_base_cross_logitadjust_real_frozenemb_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE điều chỉnh logit theo prior (τ=1) · lr 2e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.018m1 98.03% ±0.0494.12% ±0.3447.18% ±2.5089.85% ±2.2389.19% ±3.5664.29% ±0.0097.59% ±0.2093.60% ±0.4346.40% ±2.8089.76% ±2.2589.33% ±3.1537.50% ±0.00
75 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.026m1 98.22% ±0.1993.91% ±0.3058.24% ±6.4792.90% ±1.9594.70% ±0.6365.28% ±2.4197.87% ±0.2093.46% ±0.3457.20% ±6.3692.84% ±1.9794.63% ±0.6442.96% ±2.57
76 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1factorized · train thật, nhãn m1 jhu-clsp/mmBERT-base0.072m1 97.69% ±0.3693.60% ±0.4854.26% ±3.5693.47% ±2.4492.38% ±1.6464.66% ±9.8797.30% ±0.4392.96% ±0.6353.28% ±3.2993.40% ±2.4692.29% ±1.6638.36% ±14.42
77 e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 3e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.082m1 97.78% ±0.7694.11% ±0.6153.05% ±1.7594.90% ±0.8492.91% ±1.9260.83% ±6.8297.35% ±0.9393.52% ±0.5952.36% ±1.9094.85% ±0.8592.82% ±1.9435.56% ±11.76
78 e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1Qwen3-1.7B Base · 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 Qwen/Qwen3-1.7B-Base0.067m1 98.05%94.51%44.98%93.20%93.90%66.67%97.70%94.10%44.53%93.14%93.83%44.44%
79 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/visobert0.037m1 98.49%92.61%39.16%91.09%91.57%69.23%98.29%92.20%39.18%91.00%91.46%50.00%
80 e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.027m1 97.76%92.96%47.35%91.51%91.67%69.23%97.41%92.38%47.21%91.43%91.57%42.86%
81 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-base0.057m1 97.85% ±0.0993.69% ±0.8150.41% ±5.0496.61% ±1.4897.88% ±1.9583.80% ±7.6597.51% ±0.1493.14% ±0.9350.15% ±4.2696.57% ±1.5097.86% ±1.9855.56% ±9.62
82 e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_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 2e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 vinai/phobert-base-v20.206m1 97.39% ±0.4493.42% ±0.0666.36% ±13.3791.80% ±2.1091.55% ±1.6567.97% ±15.1596.90% ±0.5292.90% ±0.0765.57% ±13.5891.72% ±2.1291.45% ±1.6648.21% ±20.12
83 e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_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 cổng trộn u/v, head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.034m1 97.92% ±0.0393.40% ±0.7857.48% ±3.5891.54% ±0.8690.21% ±1.4874.77% ±12.2797.53% ±0.0692.68% ±0.8556.77% ±3.9091.46% ±0.8790.03% ±1.4550.00% ±16.67
84 e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_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 2e-05, 4 epoch · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1 vinai/phobert-base-v20.052m1 97.64% ±0.2194.14% ±0.8556.25% ±7.0693.26% ±1.5992.18% ±1.1172.38% ±13.2097.20% ±0.3193.55% ±0.8155.26% ±6.7593.19% ±1.6092.08% ±1.1341.67% ±14.43
85 e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_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 + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1 vinai/phobert-base-v20.033m1 97.57% ±0.3594.06% ±0.6055.37% ±4.4992.35% ±2.3292.53% ±1.1375.84% ±12.6597.12% ±0.4693.54% ±0.5855.01% ±4.3492.27% ±2.3492.44% ±1.1445.56% ±23.65
86 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.144m1 97.91% ±0.3293.53% ±0.0651.50% ±3.7294.61% ±2.9291.31% ±0.5564.48% ±2.0997.58% ±0.4892.88% ±0.0950.29% ±3.6494.55% ±2.9591.56% ±0.1836.94% ±3.37
87 e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.082m1 98.09% ±0.2594.12% ±0.8851.91% ±7.2291.26% ±2.0991.87% ±0.6863.18% ±14.5797.79% ±0.2593.68% ±0.9751.25% ±7.0091.18% ±2.1191.77% ±0.6946.63% ±17.38
88 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.108m1 97.81% ±0.6593.55% ±0.7857.88% ±5.4494.42% ±3.0691.82% ±1.4966.07% ±4.7297.46% ±0.8292.98% ±0.8457.03% ±5.2394.36% ±3.0891.75% ±1.4642.50% ±6.61
89 e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-0.6B-Base0.075m1 98.23% ±0.2594.44% ±0.6250.42% ±6.5591.19% ±1.8095.28% ±2.0363.43% ±2.8997.85% ±0.2893.94% ±0.6249.73% ±6.7491.11% ±1.8295.22% ±2.0543.52% ±1.60
90 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t1_tsa_m1 vinai/phobert-base-v20.070m1 97.99% ±0.1793.99% ±0.4953.51% ±3.0291.24% ±1.2293.22% ±2.3078.39% ±8.4797.63% ±0.2993.33% ±0.5352.86% ±3.2691.16% ±1.2393.13% ±2.3360.48% ±10.72
91 e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1Qwen3-0.6B Base · 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 Qwen/Qwen3-0.6B-Base0.077m1 97.26% ±0.3792.65% ±0.2745.47% ±5.0394.97% ±1.0194.78% ±0.6573.81% ±8.5896.70% ±0.4592.03% ±0.2947.40% ±7.5394.92% ±1.0294.72% ±0.6650.00% ±16.67
92 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_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-v20.093m1 98.42% ±0.3693.40% ±1.7855.82% ±1.3090.78% ±2.8491.39% ±2.6770.19% ±10.9898.22% ±0.3792.96% ±1.8554.88% ±1.3690.69% ±2.8691.29% ±2.7050.95% ±8.61
93 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_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, nhãn m1 vinai/phobert-base-v20.104m1 97.81% ±0.5593.05% ±0.4558.93% ±8.8791.52% ±1.5692.58% ±1.3381.20% ±17.0797.34% ±0.6792.37% ±0.6158.33% ±8.9591.44% ±1.5792.49% ±1.3467.20% ±29.11
94 e2_xlmr_base_cross_ce_real_frozenemb_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.044m1 97.78% ±0.4593.32% ±0.5850.47% ±7.9789.38% ±0.4389.82% ±1.9369.80% ±3.0497.43% ±0.5992.85% ±0.5649.96% ±8.4289.28% ±0.4489.73% ±1.9936.48% ±10.19
95 e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE trọng số căn nghịch tần suất · lr 2e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.019m1 97.67% ±0.5393.42% ±0.6251.70% ±5.3989.79% ±0.7990.04% ±2.2166.83% ±3.7597.26% ±0.6592.80% ±0.6150.80% ±5.5589.69% ±0.7989.92% ±2.2444.29% ±5.15
96 e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_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 điều chỉnh logit theo prior (τ=0.25) · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.049m1 98.04% ±0.6593.75% ±0.6151.80% ±0.8192.92% ±1.8091.37% ±1.3661.57% ±5.6197.72% ±0.8093.20% ±0.5451.38% ±1.1892.87% ±1.8491.27% ±1.3835.56% ±3.85
97 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_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 1e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.159m1 97.52% ±0.3192.87% ±0.8556.24% ±7.9991.80% ±1.0792.40% ±2.0867.07% ±8.8097.07% ±0.4392.51% ±0.6855.95% ±7.3791.72% ±1.0892.31% ±2.1048.21% ±9.94
98 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsa_m1 vinai/phobert-base-v20.085m1 97.85% ±0.2694.30% ±0.1556.31% ±5.0991.54% ±0.0591.79% ±1.1870.97% ±2.0397.44% ±0.2593.67% ±0.1755.62% ±5.0091.46% ±0.0591.69% ±1.2046.67% ±5.77
99 e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1 Qwen/Qwen3-Embedding-4B0.061m1 96.27% ±0.7292.02% ±0.2247.85% ±1.0498.26% ±0.9590.75% ±5.2267.92% ±31.2395.29% ±1.0591.21% ±0.0946.58% ±1.5398.24% ±0.9690.56% ±5.3951.39% ±33.39
100 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.025m1 97.51%93.40%54.98%90.57%88.10%50.00%97.08%92.91%55.84%90.48%87.95%33.33%
101 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1Qwen3-Embedding-0.6B · cross-attention (Q = câu KH, K/V = câu bot), backbone đóng băng · CE trọng số căn nghịch tần suất, label smoothing 0.05 · lr 0.0003, 40 epoch · train thật, nhãn m1 Qwen/Qwen3-Embedding-0.6B0.152m1 96.65% ±0.7290.92% ±1.2452.38% ±5.0391.40% ±1.3990.04% ±1.8358.75% ±5.3495.82% ±0.9490.11% ±1.3551.01% ±5.3391.32% ±1.4189.91% ±1.8537.48% ±5.87
102 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_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 5e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.035m1 98.17% ±0.0893.77% ±1.0253.58% ±5.4291.75% ±1.0390.91% ±3.3070.83% ±7.2297.92% ±0.1593.22% ±1.0052.68% ±5.4391.67% ±1.0490.80% ±3.3346.67% ±5.77
103 e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-base0.052m1 97.58%93.23%47.95%87.16%94.94%50.00%97.09%92.69%48.03%87.04%94.87%40.00%
104 e2_visobert_cross_ce_real_frozenemb_lr2e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 uitnlp/visobert0.027m1 97.55%91.41%41.79%93.00%90.70%58.33%97.13%91.07%41.57%92.93%90.59%16.67%
105 e2_visobert_cross_ce_real_frozenemb_lr1e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 uitnlp/visobert0.186m1 97.01%89.68%44.09%92.08%88.46%57.14%96.24%88.99%43.57%92.00%89.33%25.00%
106 e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.161m1 96.79%92.95%52.70%87.27%91.36%46.67%96.13%92.40%52.11%87.16%91.25%11.11%
107 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-0.6B0.149m1 96.28%89.39%46.81%92.63%93.42%72.73%95.49%88.19%46.61%92.55%93.33%40.00%
108 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-IDF0.207m1 96.05%89.25%57.61%91.18%94.67%80.00%95.19%88.12%56.90%91.09%94.59%50.00%
109 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-IDF0.164m1 96.24%89.57%54.97%91.18%94.67%72.73%95.43%88.46%54.14%91.09%94.59%40.00%
110 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-IDF0.280m1 96.70%89.56%46.81%90.29%95.89%76.92%96.03%88.43%46.61%90.20%95.83%57.14%
111 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-IDF0.178m1 97.00%88.25%54.77%91.18%94.67%72.73%96.34%87.00%54.26%91.09%94.59%40.00%
112 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-IDF0.267m1 96.71%90.21%46.91%90.29%95.83%81.82%96.03%89.13%45.73%90.20%95.77%60.00%
113 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-IDF0.298m1 96.57%89.77%48.25%90.29%94.67%71.43%95.87%88.66%48.58%90.20%94.59%50.00%
114 e2_dual_qwen3emb06b_concat_ce_real_frozen_m1Qwen3-Embedding-0.6B · dual encoder (câu bot và câu KH mã hoá riêng, pooling last_token), backbone đóng băng · fusion [u; v; |u−v|; u·v], head MLP 512 · CE · lr head 0.001, 30 epoch · train thật, nhãn m1 Qwen/Qwen3-Embedding-0.6B0.227m1 94.67% ±0.8789.29% ±0.7345.40% ±3.3095.08% ±0.4685.73% ±1.8892.96% ±6.1293.36% ±1.0788.43% ±0.6944.57% ±3.2195.03% ±0.4685.53% ±1.9080.56% ±17.35
115 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-IDF0.282m1 96.57%89.68%46.58%90.29%94.59%71.43%95.86%88.55%46.15%90.20%94.52%50.00%
116 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-IDF0.190m1 96.88%88.07%55.10%91.18%94.59%72.73%96.17%86.81%54.30%91.09%94.52%40.00%
117 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-IDF0.310m1 96.64%89.60%48.68%90.29%94.74%76.92%95.94%88.48%48.15%90.20%94.67%57.14%
118 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-IDF0.356m1 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/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_bamibert_cross_ce_real_frozenemb_lr2e5_m1 ––m1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_bamibert_cross_ce_real_frozenemb_lr4e5_m1 ––m1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e3_fbverifier_devtop1_familygeomean_fbf1_m1fallback_verifier · train thật, nhãn m1 –hardm1 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_m1fallback_verifier · train thật, nhãn m1 –hardm1 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_allensemble trung bình trên 23 phương pháp · nhãn v1 –0.065v1 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_allensemble chọn tham lam Caruana trên 23 phương pháp · nhãn v1 –0.070v1 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_lr4e5XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.023v1 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-v20.011v1 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_frozenembXLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE trọng số căn nghịch tần suất · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.021v1 95.95% ±0.1792.63% ±0.5953.17% ±2.2392.43% ±3.1287.13% ±2.4147.62% ±17.1795.16% ±0.4092.05% ±0.6851.41% ±1.5992.32% ±3.1686.95% ±2.4437.78% ±20.37
6 e3_prior_em_mmbert_bottypeEM điều chỉnh prior trên e2_mmbert_base_cross_ce_real_frozenemb_lr2e5, phạm vi theo loại câu bot (transductive) · nhãn v1 –0.082v1 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_globalEM điều chỉnh prior trên e2_mmbert_base_cross_ce_real_frozenemb_lr2e5, phạm vi toàn bộ (transductive) · nhãn v1 –0.082v1 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_lr2e5mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 jhu-clsp/mmBERT-base0.061v1 96.18% ±0.3093.86% ±0.9250.93% ±3.2893.05% ±3.4088.45% ±2.4465.08% ±7.2795.71% ±0.4193.52% ±0.8849.18% ±3.4292.95% ±3.4488.28% ±2.4758.89% ±8.39
9 e2_xlmr_base_cross_ce_real_frozenembXLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.035v1 95.85% ±0.4091.77% ±0.6756.89% ±0.5892.70% ±0.7087.31% ±1.6854.17% ±19.0995.07% ±0.3891.11% ±0.7455.39% ±0.5892.60% ±0.7187.12% ±1.7045.08% ±19.55
10 e2_xlmr_base_cross_logitadjust_real_frozenembXLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE điều chỉnh logit theo prior (τ=1) · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.035v1 96.38% ±0.3792.66% ±0.6250.92% ±2.6093.59% ±2.0690.11% ±2.1935.73% ±2.1595.80% ±0.3892.08% ±0.6949.04% ±2.1093.50% ±2.0989.96% ±2.2327.86% ±2.58
11 e2_qwen3_06b_base_cross_ce_real_loraQwen3-0.6B Base · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn v1 Qwen/Qwen3-0.6B-Base0.098v1 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_allensemble stacking LR trên 23 phương pháp · nhãn v1 –0.088v1 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_loraQwen3-1.7B Base · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật, nhãn v1 Qwen/Qwen3-1.7B-Base0.083v1 95.93% ±0.1092.55% ±1.1150.23% ±2.7295.71% ±1.3591.01% ±1.3458.33% ±7.2295.19% ±0.0792.16% ±1.2248.72% ±2.7995.65% ±1.3690.88% ±1.3651.43% ±9.90
14 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 vinai/phobert-base-v20.022v1 95.73% ±0.2892.72% ±0.4155.00% ±0.9794.07% ±3.2387.83% ±2.9477.78% ±19.2594.91% ±0.3692.24% ±0.4853.32% ±1.2093.99% ±3.2887.65% ±2.9866.67% ±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-base0.088v1 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_lr4e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 intfloat/multilingual-e5-base0.028v1 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_frozenembXLM-R base · chỉ câu KH · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.076v1 95.11% ±0.1891.66% ±1.2556.10% ±4.0991.48% ±0.7190.53% ±1.4761.11% ±9.6294.24% ±0.3690.97% ±1.4254.13% ±3.9691.37% ±0.7290.38% ±1.4947.78% ±13.47
18 e2_me5_base_cross_ce_real_frozenemb_lr2e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 intfloat/multilingual-e5-base0.056v1 96.41% ±0.2491.29% ±0.8855.12% ±1.6190.28% ±1.4088.66% ±2.7855.00% ±18.0395.85% ±0.3390.77% ±0.8752.97% ±1.6790.15% ±1.4288.49% ±2.8241.67% ±22.05
19 e2_gemma3_270m_cross_ce_real_loragemma-3-270m · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn v1 unsloth/gemma-3-270m0.116v1 95.90% ±0.5092.19% ±0.8350.12% ±3.3792.12% ±1.9887.27% ±1.4442.59% ±8.4995.15% ±0.4191.53% ±0.7848.70% ±3.4092.45% ±1.9987.53% ±1.0934.17% ±8.04
20 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozenQwen3-Embedding-0.6B · cross-attention (Q = câu KH, K/V = câu bot), backbone đóng băng · CE trọng số căn nghịch tần suất, label smoothing 0.05 · lr 0.0003, 40 epoch · train thật, nhãn v1 Qwen/Qwen3-Embedding-0.6B0.114v1 95.24% ±0.1691.30% ±0.2349.00% ±1.1889.20% ±2.9491.10% ±2.4851.43% ±9.9094.20% ±0.2090.81% ±0.2547.21% ±1.4989.05% ±2.9790.97% ±2.5244.44% ±9.62
21 e2_mmbert_base_cross_ce_real_frozenemb_lr4e5mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 jhu-clsp/mmBERT-base0.075v1 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_lr1e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 vinai/phobert-base-v20.026v1 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-base0.123v1 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_lr1e5mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 jhu-clsp/mmBERT-base0.120v1 95.81% ±0.0591.40% ±0.8455.08% ±1.0195.75% ±1.9686.36% ±4.2936.43% ±9.0995.35% ±0.0990.98% ±0.9453.93% ±1.5795.69% ±1.9986.72% ±3.5527.78% ±7.86
25 e2_visobert_cross_ce_real_frozenemb_lr2e5ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 uitnlp/visobert0.093v1 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_lr4e5ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 uitnlp/visobert0.066v1 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-IDF0.123v1 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-IDF0.256v1 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-IDF0.125v1 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_lr1e5ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 uitnlp/visobert0.232v1 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-IDF0.234v1 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/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_mdeberta_cross_ce_real_frozenemb_lr2e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 ––v1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_mdeberta_cross_ce_real_frozenemb_lr2e5_fp32mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 ––v1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_mdeberta_cross_ce_real_frozenemb_lr4e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 microsoft/mdeberta-v3-baseargmaxv1 N/A23.81%N/AN/AN/AN/AN/A26.63%N/AN/AN/AN/A
– e2_videberta_cross_ce_real_frozenemb_lr1e5ViDeBERTa · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 Fsoft-AIC/videberta-baseargmaxv1 N/A23.81%N/AN/AN/AN/AN/A26.63%N/AN/AN/AN/A
– e2_videberta_cross_ce_real_frozenemb_lr2e5ViDeBERTa · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 Fsoft-AIC/videberta-baseargmaxv1 N/A23.81%N/AN/AN/AN/AN/A26.63%N/AN/AN/AN/A
– e2_videberta_cross_ce_real_frozenemb_lr2e5_fp32 ––v1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_videberta_cross_ce_real_frozenemb_lr4e5ViDeBERTa · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 Fsoft-AIC/videberta-baseargmaxv1 N/A23.81%N/AN/AN/AN/AN/A26.63%N/AN/AN/AN/A
1 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_lv21mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v2_1 jhu-clsp/mmBERT-base0.076v2_1 95.67% ±0.4190.20% ±0.6854.63% ±4.6297.52% ±1.6790.50% ±2.3060.19% ±13.1395.03% ±0.5790.12% ±0.6253.83% ±3.8397.48% ±1.7090.35% ±2.3450.00% ±16.67
2 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_lv21PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v2_1 vinai/phobert-base-v20.033v2_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

STTModel_NameBackboneThresholdData clean Affirm_Recclean Negate_Recclean Fallback_Recclean Callback_Recclean Relation_Recclean Repeat_Recreply-unseen Affirm_Recreply-unseen Negate_Recreply-unseen Fallback_Recreply-unseen Callback_Recreply-unseen Relation_Recreply-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_m2 –0.051m2 95.79%89.49%99.20%98.97%89.53%75.00%94.76%88.76%100.00%98.96%90.48%50.00%
2 e3_ensemble_devtop1_familygeomean_tempdev_m2 –0.051m2 96.05%89.90%99.20%100.00%89.53%75.00%95.03%89.21%100.00%100.00%90.48%50.00%
3 e3_ensemble_devtop1_familygeomean_temphard_m2 –0.051m2 96.05%89.90%99.20%100.00%89.53%75.00%95.03%89.21%100.00%100.00%90.48%50.00%
4 e3_ensemble_devtop1_familymean_m2 –0.112m2 96.12%89.70%98.40%100.00%90.70%83.33%95.11%88.99%99.15%100.00%91.67%66.67%
5 e3_ensemble_sweep_k2_m2 –0.110m2 97.19%92.32%96.00%98.97%88.37%83.33%96.51%91.91%96.58%98.96%89.29%66.67%
6 e3_ensemble_sweep_k3_m2 –0.051m2 96.52%91.11%96.80%100.00%88.37%75.00%95.64%90.56%97.44%100.00%89.29%50.00%
7 e3_ensemble_sweep_k4_m2 –0.058m2 96.72%90.91%96.80%100.00%89.53%75.00%95.90%90.34%97.44%100.00%90.48%50.00%
8 e3_ensemble_sweep_k5_m2 –0.111m2 97.32%93.13%97.60%100.00%89.53%83.33%96.68%92.81%98.29%100.00%90.48%66.67%
9 e3_ensemble_sweep_k6_m2 –0.087m2 97.12%92.32%97.60%100.00%89.53%83.33%96.42%91.91%98.29%100.00%90.48%66.67%
10 e3_ensemble_sweep_k7_m2 –0.087m2 97.12%92.32%97.60%100.00%89.53%83.33%96.42%91.91%98.29%100.00%90.48%66.67%
11 e3_ensemble_sweep_k8_m2 –0.087m2 97.12%92.32%97.60%100.00%89.53%83.33%96.42%91.91%98.29%100.00%90.48%66.67%
12 e3_ensemble_top_geomean_m2 –0.073m2 96.25%91.11%97.60%100.00%91.86%83.33%95.29%90.56%98.29%100.00%92.86%66.67%
13 e3_ensemble_top_mean_m2 –0.073m2 94.98%88.48%98.40%100.00%91.86%83.33%93.72%87.87%99.15%100.00%92.86%66.67%
14 e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m2 unsloth/gemma-3-1b-pt0.064m2 94.36% ±1.6089.56% ±1.6896.00% ±0.8097.94% ±1.7986.82% ±2.9366.67% ±8.3392.99% ±1.8888.84% ±1.8796.58% ±0.8597.92% ±1.8087.70% ±3.0033.33% ±16.67
15 e2_me5_base_cross_ce_real_frozenemb_lr4e5_m2 intfloat/multilingual-e5-base0.017m2 92.64%89.70%96.00%100.00%82.56%75.00%90.66%88.76%95.73%100.00%83.33%50.00%
16 e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m2 jhu-clsp/mmBERT-base0.155m2 96.99%89.29%88.00%95.88%83.72%83.33%96.16%88.54%88.03%95.83%84.52%66.67%
17 e2_qwen3_17b_base_genlabel_ce_real_lora_m2 Qwen/Qwen3-1.7B-Base0.081m2 95.05%90.30%93.60%98.97%88.37%91.67%93.72%89.66%93.16%98.96%89.29%83.33%
18 e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m2 Qwen/Qwen3-1.7B-Base0.105m2 95.12% ±0.7489.90% ±0.9395.47% ±0.4698.28% ±1.5791.09% ±1.3491.67% ±0.0093.89% ±0.9289.14% ±1.1395.16% ±0.4998.26% ±1.5991.67% ±1.1983.33% ±0.00
19 e3_ensemble_top_greedy_m2 –0.113m2 95.65%90.51%96.80%100.00%90.70%91.67%94.50%89.89%97.44%100.00%91.67%83.33%
20 e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m2 intfloat/multilingual-e5-large-instruct0.017m2 95.30% ±0.6291.04% ±2.4796.27% ±0.4698.28% ±1.1985.27% ±3.3683.33% ±0.0094.10% ±0.7890.49% ±2.5396.58% ±0.8598.26% ±1.2086.11% ±3.4466.67% ±0.00
21 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m2 jhu-clsp/mmBERT-base0.094m2 94.58% ±1.7089.63% ±2.1695.73% ±3.2398.28% ±0.6084.88% ±2.0172.22% ±17.3593.08% ±2.1689.14% ±2.2696.01% ±3.2498.26% ±0.6085.71% ±2.0644.44% ±34.69
22 e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m2 Qwen/Qwen3-4B-Base0.056m2 96.05% ±0.4688.22% ±0.4796.27% ±0.4698.63% ±0.6092.64% ±1.7880.56% ±4.8195.17% ±0.5587.49% ±0.2696.01% ±0.4998.61% ±0.6092.86% ±2.0661.11% ±9.62
23 e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m2 unsloth/gemma-3-1b-pt0.045m2 94.92% ±0.1988.79% ±0.4396.40% ±0.5796.91% ±1.4687.79% ±5.7675.00% ±0.0093.63% ±0.2587.98% ±0.4897.01% ±0.6096.88% ±1.4788.69% ±5.8950.00% ±0.00
24 e2_gemma3_1b_pt_cross_ce_real_lora_r32_m2 unsloth/gemma-3-1b-pt0.045m2 94.56% ±1.1486.53% ±1.3595.47% ±0.9296.56% ±1.1987.98% ±3.7480.56% ±4.8193.22% ±1.4785.47% ±1.5096.01% ±0.9996.53% ±1.2088.89% ±3.8361.11% ±9.62
25 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m2 jhu-clsp/mmBERT-base0.052m2 94.63% ±0.4588.82% ±2.6992.80% ±1.3995.88% ±2.7385.27% ±5.7491.67% ±8.3393.72% ±0.4087.94% ±2.9092.88% ±1.3195.83% ±2.7686.11% ±5.8783.33% ±16.67
26 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m2 vinai/phobert-base-v20.008m2 95.72%89.29%95.20%97.94%87.21%16.67%94.76%88.76%96.58%97.92%88.10%16.67%
27 e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m2 Qwen/Qwen3-4B-Base0.162m2 95.79%85.05%97.60%100.00%91.86%100.00%94.85%83.82%97.44%100.00%92.86%100.00%
28 e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m2 Qwen/Qwen3-8B-Base0.170m2 96.86%90.51%94.40%98.97%91.86%91.67%96.07%90.11%94.02%98.96%92.86%83.33%
29 e3_ensemble_seed6_kd_a05_t1_m2 –0.067m2 96.12%89.49%96.80%100.00%90.70%75.00%95.11%88.99%97.44%100.00%91.67%50.00%
30 e2_me5_base_cross_ce_real_frozenemb_lr2e5_m2 intfloat/multilingual-e5-base0.031m2 94.29% ±0.8489.97% ±2.5396.53% ±1.8599.31% ±0.6086.82% ±1.7863.89% ±4.8192.79% ±1.0589.21% ±2.7096.30% ±1.9799.31% ±0.6087.70% ±1.8227.78% ±9.62
31 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m2 jhu-clsp/mmBERT-base0.061m2 94.56% ±0.7288.35% ±3.2195.47% ±1.6797.59% ±1.1985.27% ±4.8477.78% ±4.8193.37% ±0.6987.57% ±3.4495.73% ±1.7197.57% ±1.2086.11% ±4.9655.56% ±9.62
32 e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m2 Qwen/Qwen3-4B-Base0.131m2 96.77% ±1.0289.29% ±1.2395.73% ±2.0198.97% ±1.0391.47% ±1.7886.11% ±4.8196.02% ±1.2488.61% ±1.5095.44% ±2.1598.96% ±1.0492.46% ±1.8272.22% ±9.62
33 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.043m2 95.03% ±0.4988.01% ±0.6294.13% ±2.5798.97% ±1.0388.76% ±4.7072.22% ±9.6293.72% ±0.5487.57% ±0.4794.02% ±2.2698.96% ±1.0489.68% ±4.8144.44% ±19.25
34 e2_gemma3_1b_pt_cross_ce_real_lora_m2 unsloth/gemma-3-1b-pt0.053m2 95.36% ±0.4988.35% ±1.5395.20% ±0.0096.56% ±1.1986.82% ±3.3675.00% ±0.0094.27% ±0.6287.72% ±1.5396.01% ±0.4996.53% ±1.2087.70% ±3.4450.00% ±0.00
35 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m2 vinai/phobert-base-v20.085m2 96.32% ±0.4788.22% ±1.9893.60% ±0.80100.00% ±0.0089.15% ±1.7880.56% ±12.7395.52% ±0.5687.42% ±2.1494.02% ±0.85100.00% ±0.0090.08% ±1.8261.11% ±25.46
36 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m2 jhu-clsp/mmBERT-base0.082m2 95.38% ±0.5888.22% ±3.2593.60% ±3.6796.91% ±1.0384.11% ±3.5580.56% ±4.8194.15% ±0.8387.42% ±3.3393.73% ±3.5696.88% ±1.0484.92% ±3.6461.11% ±9.62
37 e2_qwen3_17b_base_cross_ce_real_lora_m2 Qwen/Qwen3-1.7B-Base0.101m2 94.80% ±0.7988.48% ±0.8193.07% ±2.4498.28% ±1.5791.09% ±1.7880.56% ±9.6293.57% ±0.9687.64% ±0.9093.45% ±2.6198.26% ±1.5992.06% ±1.8261.11% ±19.25
38 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m2 vinai/phobert-base-v20.068m2 95.81% ±0.5787.95% ±2.2095.87% ±1.1899.83% ±0.4290.50% ±1.5576.39% ±9.7494.76% ±0.7787.38% ±2.2996.30% ±1.2999.83% ±0.4391.47% ±1.5852.78% ±19.48
39 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m2 vinai/phobert-base-v20.052m2 95.18% ±0.8088.55% ±4.0595.47% ±1.2298.97% ±1.0389.15% ±2.9375.00% ±14.4394.01% ±0.9387.94% ±4.2995.44% ±1.7898.96% ±1.0489.68% ±2.4850.00% ±28.87
40 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m2 vinai/phobert-base-v20.058m2 95.23% ±0.1088.22% ±0.8494.40% ±0.8098.28% ±2.1588.37% ±4.6580.56% ±12.7394.01% ±0.1387.64% ±1.0394.87% ±0.0098.26% ±2.1789.29% ±4.7661.11% ±25.46
41 e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m2 unsloth/gemma-3-1b-pt0.056m2 95.43% ±0.5186.80% ±1.9194.40% ±0.8097.25% ±0.6086.05% ±4.1980.56% ±4.8194.33% ±0.6885.92% ±1.8795.16% ±1.7897.22% ±0.6086.90% ±4.2961.11% ±9.62
42 e2_gemma3_270m_cross_ce_real_lora_m2 unsloth/gemma-3-270m0.098m2 94.56% ±0.5685.05% ±1.7393.07% ±1.2296.56% ±0.6081.78% ±6.4083.33% ±8.3393.25% ±0.8883.82% ±1.9293.73% ±0.9996.53% ±0.6082.54% ±6.5666.67% ±16.67
43 e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m2 intfloat/multilingual-e5-large-instruct0.029m2 95.63% ±0.6090.44% ±0.9194.67% ±1.6798.63% ±0.6085.27% ±1.7872.22% ±9.6294.62% ±0.7489.89% ±1.0395.44% ±1.3198.61% ±0.6086.11% ±1.8244.44% ±19.25
44 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m2 jhu-clsp/mmBERT-base0.040m2 94.29% ±1.0187.95% ±2.4195.47% ±0.9296.91% ±2.7383.33% ±5.8572.22% ±4.8192.84% ±1.2187.12% ±2.4095.44% ±0.4996.88% ±2.7684.13% ±5.9944.44% ±9.62
45 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m2 jhu-clsp/mmBERT-base0.057m2 94.80% ±1.3788.82% ±0.3195.73% ±1.2297.25% ±1.5786.05% ±2.0183.33% ±0.0093.78% ±1.0687.94% ±0.4796.01% ±1.3197.22% ±1.5986.90% ±2.0666.67% ±0.00
46 e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m2 jhu-clsp/mmBERT-base0.070m2 95.52%90.51%94.40%97.94%83.72%91.67%94.59%89.89%94.87%97.92%84.52%83.33%
47 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m2 jhu-clsp/mmBERT-base0.047m2 94.76% ±0.8289.16% ±0.7193.07% ±1.2296.91% ±2.0686.05% ±3.0875.00% ±8.3393.48% ±1.0888.46% ±0.6993.45% ±1.3196.88% ±2.0886.90% ±3.1550.00% ±16.67
48 e2_xlmr_base_khonly_ce_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.012m2 94.05%86.26%96.00%100.00%88.37%66.67%92.41%85.62%96.58%100.00%89.29%33.33%
49 e3_ensemble_seed_kd_a05_t1_m2 –0.078m2 95.85%89.29%96.80%100.00%91.86%75.00%94.85%88.99%97.44%100.00%92.86%50.00%
50 e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m2 unsloth/gemma-3-1b-pt0.042m2 94.60% ±0.8287.68% ±1.2194.93% ±1.6796.91% ±1.7983.72% ±1.1680.56% ±9.6293.48% ±0.9086.82% ±1.3595.44% ±1.7896.88% ±1.8084.52% ±1.1961.11% ±19.25
51 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m2 vinai/phobert-base-v20.024m2 94.89% ±0.1986.67% ±1.3296.00% ±0.80100.00% ±0.0087.98% ±6.8172.22% ±9.6293.57% ±0.2286.14% ±1.3596.01% ±0.49100.00% ±0.0088.89% ±6.9844.44% ±19.25
52 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m2 vinai/phobert-base-v20.031m2 95.79% ±0.5486.26% ±2.7893.60% ±2.1298.97% ±1.7986.43% ±6.4075.00% ±14.4394.71% ±0.6585.47% ±2.8394.02% ±2.2698.96% ±1.8087.30% ±6.5650.00% ±28.87
53 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t2_tsa_m2 vinai/phobert-base-v20.070m2 95.54% ±0.5288.08% ±1.8093.60% ±2.4099.31% ±0.6089.15% ±3.7488.89% ±4.8194.27% ±0.6887.64% ±2.0094.59% ±2.1599.31% ±0.6089.68% ±3.6477.78% ±9.62
54 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m2 vinai/phobert-base-v20.031m2 95.41% ±1.0088.35% ±0.6596.00% ±0.8098.63% ±1.1986.05% ±3.4975.00% ±8.3394.21% ±1.3587.87% ±0.5996.30% ±1.3198.61% ±1.2086.90% ±3.5750.00% ±16.67
55 e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m2 unsloth/gemma-3-1b-pt0.096m2 95.34% ±1.0087.81% ±1.9292.00% ±0.0097.25% ±0.6088.37% ±3.0855.56% ±26.7994.27% ±1.3187.12% ±2.3692.31% ±0.0097.22% ±0.6089.29% ±3.1538.89% ±9.62
56 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m2 intfloat/multilingual-e5-large-instruct0.038m2 94.83% ±0.8187.21% ±2.3796.00% ±0.8098.28% ±1.5787.60% ±3.5580.56% ±4.8193.43% ±0.9986.44% ±2.4797.15% ±0.4998.26% ±1.5988.49% ±3.6461.11% ±9.62
57 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.017m2 95.72% ±0.1288.96% ±1.1593.87% ±1.8598.63% ±0.6087.98% ±2.4258.33% ±0.0094.65% ±0.2288.31% ±1.4094.87% ±2.2698.61% ±0.6088.89% ±2.4816.67% ±0.00
58 e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m2 vinai/phobert-base-v20.026m2 96.16% ±0.5889.83% ±1.6592.00% ±1.3999.31% ±0.6087.21% ±1.1666.67% ±8.3395.23% ±0.6289.29% ±1.6993.16% ±1.4899.31% ±0.6088.10% ±1.1933.33% ±16.67
59 e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.065m2 95.34% ±0.6387.61% ±1.5494.40% ±1.6098.28% ±1.5787.98% ±0.6777.78% ±4.8194.21% ±0.8786.97% ±1.6294.59% ±1.3198.26% ±1.5988.89% ±0.6955.56% ±9.62
60 e2_xlmr_base_cross_logitadjust_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.018m2 94.11% ±0.6684.65% ±1.8095.20% ±1.6099.31% ±0.6088.76% ±2.4275.00% ±0.0092.52% ±0.9083.75% ±1.7595.73% ±2.2699.31% ±0.6089.29% ±2.0650.00% ±0.00
61 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m2 intfloat/multilingual-e5-large-instruct0.026m2 95.41% ±0.8092.73% ±0.7094.13% ±1.6797.94% ±1.7987.98% ±1.7883.33% ±0.0094.21% ±1.1092.28% ±0.7294.02% ±2.2697.92% ±1.8088.89% ±1.8266.67% ±0.00
62 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m2 jhu-clsp/mmBERT-base0.072m2 95.61% ±0.9389.76% ±1.7290.93% ±2.5796.56% ±2.1583.33% ±7.1075.00% ±8.3394.44% ±1.1389.14% ±1.8391.17% ±2.7596.53% ±2.1784.13% ±7.2750.00% ±16.67
63 e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m2 jhu-clsp/mmBERT-base0.082m2 95.12% ±0.6189.43% ±1.7294.67% ±1.6795.53% ±2.5983.72% ±7.0777.78% ±9.6293.89% ±0.7588.61% ±1.8395.16% ±1.7895.49% ±2.6284.52% ±7.2455.56% ±19.25
64 e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m2 Qwen/Qwen3-1.7B-Base0.067m2 93.58%86.46%93.60%98.97%87.21%83.33%92.06%85.39%94.02%98.96%88.10%66.67%
65 e2_visobert_cross_ce_real_frozenemb_lr4e5_m2 uitnlp/visobert0.037m2 91.24%85.25%94.40%94.85%84.88%75.00%89.70%84.04%94.87%94.79%85.71%50.00%
66 e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m2 FacebookAI/xlm-roberta-base0.027m2 92.91%89.49%90.40%100.00%87.21%75.00%91.27%88.76%91.45%100.00%88.10%50.00%
67 e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m2 jhu-clsp/mmBERT-base0.057m2 94.72% ±0.5788.82% ±2.7696.00% ±1.6097.25% ±0.6085.27% ±2.9366.67% ±8.3393.66% ±0.4388.16% ±2.9796.58% ±1.7197.22% ±0.6086.11% ±3.0033.33% ±16.67
68 e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.034m2 95.96% ±0.2888.82% ±0.6292.53% ±1.22100.00% ±0.0089.92% ±0.6775.00% ±14.4394.85% ±0.4088.24% ±0.9192.88% ±1.31100.00% ±0.0090.08% ±1.3750.00% ±28.87
69 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m2 jhu-clsp/mmBERT-base0.144m2 94.58% ±0.5889.49% ±0.8893.60% ±0.8094.50% ±2.3886.82% ±1.7875.00% ±8.3393.11% ±0.6888.84% ±0.9493.73% ±1.3194.44% ±2.4187.70% ±1.8250.00% ±16.67
70 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m2 jhu-clsp/mmBERT-base0.108m2 95.99% ±0.8189.97% ±1.1593.60% ±2.1296.91% ±1.0384.11% ±3.7480.56% ±4.8194.94% ±0.9889.36% ±1.1594.02% ±2.2696.88% ±1.0484.92% ±3.8361.11% ±9.62
71 e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m2 Qwen/Qwen3-0.6B-Base0.075m2 94.58% ±1.2987.14% ±2.1096.00% ±1.3999.31% ±0.6090.31% ±1.7886.11% ±4.8193.31% ±1.7286.14% ±2.3496.01% ±1.3199.31% ±0.6091.27% ±1.8272.22% ±9.62
72 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t1_tsa_m2 vinai/phobert-base-v20.070m2 95.27% ±0.2087.74% ±1.9295.47% ±1.6799.66% ±0.6087.21% ±3.4986.11% ±9.6293.95% ±0.2287.19% ±2.1495.73% ±1.4899.65% ±0.6088.10% ±3.5772.22% ±19.25
73 e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m2 Qwen/Qwen3-0.6B-Base0.077m2 93.00% ±0.8786.67% ±3.5292.27% ±3.2397.25% ±1.1988.76% ±2.9377.78% ±9.6292.20% ±1.9485.69% ±3.7992.59% ±3.4597.22% ±1.2089.68% ±3.0055.56% ±19.25
74 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m2 vinai/phobert-base-v20.104m2 95.90% ±1.0989.29% ±1.8093.33% ±3.2399.66% ±0.6084.50% ±4.0877.78% ±9.6294.85% ±1.3788.76% ±1.7892.88% ±3.4599.65% ±0.6085.32% ±4.1855.56% ±19.25
75 e2_xlmr_base_cross_ce_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.044m2 94.43% ±0.9686.73% ±4.0791.73% ±3.2398.28% ±1.1988.76% ±0.6769.44% ±12.7392.96% ±1.2085.99% ±4.5092.59% ±3.2498.26% ±1.2089.68% ±0.6938.89% ±25.46
76 e2_xlmr_base_cross_wcesqrt_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.019m2 94.49% ±1.0486.20% ±1.1894.40% ±2.8899.31% ±0.6087.98% ±2.9383.33% ±8.3392.96% ±1.2785.17% ±1.3795.16% ±2.1599.31% ±0.6088.89% ±3.0066.67% ±16.67
77 e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.049m2 95.23% ±0.2387.21% ±1.8295.20% ±0.8098.63% ±1.5787.60% ±2.4277.78% ±9.6294.07% ±0.3186.59% ±2.0495.44% ±0.9998.61% ±1.5988.49% ±2.4855.56% ±19.25
78 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m2 vinai/phobert-base-v20.159m2 95.76% ±0.6887.68% ±2.8692.80% ±2.1299.66% ±0.6085.66% ±5.9761.11% ±31.5594.76% ±1.0186.97% ±3.0293.45% ±2.4799.65% ±0.6086.51% ±6.1150.00% ±16.67
79 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsa_m2 vinai/phobert-base-v20.085m2 95.74% ±0.8788.42% ±0.7194.13% ±2.44100.00% ±0.0088.76% ±3.5580.56% ±12.7394.56% ±1.0987.94% ±0.6994.30% ±2.15100.00% ±0.0089.68% ±3.6461.11% ±25.46
80 e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m2 Qwen/Qwen3-Embedding-4B0.061m2 94.15% ±0.5283.84% ±2.2992.80% ±3.3989.69% ±8.7579.65% ±4.1183.33% ±11.7992.50% ±0.6282.70% ±2.2292.31% ±3.6389.58% ±8.8479.76% ±3.3766.67% ±23.57
81 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m2 vinai/phobert-base-v20.025m2 96.39%88.69%90.40%98.97%82.56%16.67%95.38%88.31%91.45%98.96%83.33%16.67%
82 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m2 Qwen/Qwen3-Embedding-0.6B0.152m2 94.58% ±0.5486.20% ±3.1186.67% ±0.4694.85% ±0.0080.23% ±4.0386.11% ±4.8193.11% ±0.7085.02% ±3.3886.04% ±0.4994.79% ±0.0080.95% ±4.1272.22% ±9.62
83 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m2 vinai/phobert-base-v20.035m2 95.23% ±1.0788.48% ±2.3395.47% ±1.2298.97% ±1.0388.37% ±3.0877.78% ±4.8194.01% ±1.2787.72% ±2.7095.73% ±1.4898.96% ±1.0489.29% ±3.1555.56% ±9.62
84 e2_me5_base_cross_ce_real_frozenemb_lr1e5_m2 intfloat/multilingual-e5-base0.052m2 94.05%88.69%92.00%97.94%86.05%25.00%92.58%87.87%92.31%97.92%86.90%33.33%
85 e2_visobert_cross_ce_real_frozenemb_lr2e5_m2 uitnlp/visobert0.027m2 92.64%82.63%91.20%95.88%87.21%58.33%91.19%81.12%92.31%95.83%88.10%16.67%
86 e2_visobert_cross_ce_real_frozenemb_lr1e5_m2 uitnlp/visobert0.186m2 92.71%84.85%87.20%95.88%76.74%66.67%91.10%84.04%87.18%95.83%76.19%33.33%
87 e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m2 FacebookAI/xlm-roberta-base0.161m2 94.18%86.46%91.20%97.94%83.72%58.33%92.50%85.84%92.31%97.92%84.52%16.67%
88 e1_tfidf_hybrid_qwen3emb_lr_m2 Qwen/Qwen3-Embedding-0.6B0.149m2 94.65%81.82%85.60%90.72%79.07%66.67%93.54%80.45%85.47%90.62%79.76%33.33%
89 e1_tfidf_svmcal_ce_real_base_m2 TF-IDF0.178m2 94.98%87.88%84.80%94.85%81.40%66.67%93.72%86.97%84.62%94.79%82.14%33.33%
90 e2_dual_qwen3emb06b_concat_ce_real_frozen_m2 Qwen/Qwen3-Embedding-0.6B0.227m2 93.40% ±1.4479.66% ±0.6287.20% ±0.8092.78% ±3.0969.38% ±7.1066.67% ±8.3391.80% ±1.7578.20% ±0.5986.89% ±0.4992.71% ±3.1269.84% ±7.2733.33% ±16.67
– e3_fbverifier_devtop1_familygeomean_fbf1_m2 –hardm2 98.19%95.35%87.20%100.00%90.70%91.67%97.82%95.28%88.03%100.00%91.67%83.33%
– e3_fbverifier_devtop1_familygeomean_macrof1_m2 –hardm2 98.19%95.35%87.20%100.00%90.70%91.67%97.82%95.28%88.03%100.00%91.67%83.33%
1 e3_ensemble_devtop1_familygeomean_m1ensemble trung bình nhân theo nhóm trên 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051m1 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 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051m1 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 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051m1 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_m1ensemble trung bình đều theo nhóm trên 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.112m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 2 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.110m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 3 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 4 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.058m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.111m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.087m1 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 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.087m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.087m1 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_m1ensemble trung bình nhân trên 5 phương pháp · nhãn m1 –0.073m1 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_m1ensemble trung bình trên 5 phương pháp · nhãn m1 –0.073m1 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_m1gemma-3-1b-pt · 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 unsloth/gemma-3-1b-pt0.064m1 93.93% ±1.5586.64% ±1.6294.62% ±0.7797.28% ±2.1284.81% ±4.2166.67% ±8.3392.44% ±1.8285.64% ±1.7995.08% ±0.8297.25% ±2.1585.61% ±4.3033.33% ±16.67
15 e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-base0.017m1 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_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.155m1 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_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-1.7B-Base0.081m1 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_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-1.7B-Base0.105m1 94.48% ±0.8086.77% ±0.7893.33% ±0.4498.30% ±1.5689.26% ±1.7091.67% ±0.0093.07% ±1.0085.79% ±1.2592.90% ±0.4798.28% ±1.5789.77% ±1.9783.33% ±0.00
19 e3_ensemble_top_greedy_m1ensemble chọn tham lam Caruana trên 5 phương pháp · nhãn m1 –0.113m1 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_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 3e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.017m1 94.85% ±0.5987.48% ±2.6594.36% ±0.4497.62% ±0.5984.44% ±1.9283.33% ±0.0093.53% ±0.7586.57% ±2.7294.54% ±0.9597.59% ±0.6085.23% ±1.9766.67% ±0.00
21 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m1 jhu-clsp/mmBERT-base0.094m1 94.12% ±1.7286.32% ±2.2294.10% ±3.5597.28% ±0.5983.70% ±1.7072.22% ±17.3592.50% ±2.1985.50% ±2.3594.26% ±3.5797.25% ±0.6084.47% ±1.7444.44% ±34.69
22 e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1factorized · train thật + synth ×1, nhãn m1 jhu-clsp/mmBERT-base0.108m1 92.60% ±1.3784.65% ±3.2892.56% ±5.5696.26% ±0.5982.59% ±5.2583.33% ±14.4391.49% ±1.3983.51% ±3.6392.08% ±5.9396.22% ±0.6083.33% ±5.3766.67% ±28.87
23 e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-4B-Base0.056m1 95.54% ±0.4884.59% ±0.6795.38% ±0.7798.64% ±0.5991.11% ±1.1180.56% ±4.8194.54% ±0.5583.44% ±0.4495.08% ±0.8298.63% ±0.6091.29% ±1.3161.11% ±9.62
24 e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 5 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.045m1 94.43% ±0.2885.45% ±0.1495.00% ±0.5495.92% ±1.4486.67% ±6.2975.00% ±0.0093.02% ±0.2484.33% ±0.1595.49% ±0.5895.88% ±1.4687.50% ±6.4350.00% ±0.00
25 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-v20.033m1 94.01% ±0.5784.07% ±1.0691.79% ±1.6098.64% ±0.5986.30% ±2.5766.67% ±8.3392.76% ±0.8783.08% ±1.2591.80% ±1.4298.63% ±0.6087.12% ±2.6233.33% ±16.67
26 e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=32 · CE · lr 0.0001, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.045m1 93.93% ±1.1183.43% ±1.5394.10% ±0.8995.58% ±1.1888.52% ±3.5780.56% ±4.8192.39% ±1.5082.16% ±1.6194.54% ±0.9595.53% ±1.1989.39% ±3.6561.11% ±9.62
27 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 6 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.052m1 94.21% ±0.4785.36% ±2.8491.03% ±1.6095.24% ±2.9583.70% ±5.4891.67% ±8.3393.19% ±0.3884.15% ±3.0590.98% ±1.4295.19% ±2.9884.47% ±5.6183.33% ±16.67
28 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-base0.051m1 94.96% ±0.7386.64% ±2.3591.79% ±2.9195.92% ±1.0284.81% ±2.8086.11% ±4.8193.76% ±0.6985.64% ±2.6191.80% ±2.8495.88% ±1.0385.61% ±2.8672.22% ±9.62
29 e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 jhu-clsp/mmBERT-base0.118m1 95.05% ±1.0384.26% ±2.2592.82% ±3.7995.92% ±1.0285.56% ±2.9483.33% ±8.3393.76% ±1.3783.16% ±2.3792.90% ±3.8795.88% ±1.0386.36% ±3.0166.67% ±16.67
30 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1PhoBERT-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 m1 vinai/phobert-base-v20.008m1 95.03%85.74%93.08%96.94%85.56%16.67%93.88%84.86%94.26%96.91%86.36%16.67%
31 e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1 Qwen/Qwen3-4B-Base0.162m1 95.29%81.12%96.15%100.00%92.22%100.00%94.22%79.53%95.90%100.00%93.18%100.00%
32 e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1 Qwen/Qwen3-8B-Base0.170m1 96.36%87.09%92.31%98.98%92.22%91.67%95.43%86.35%91.80%98.97%93.18%83.33%
33 e3_ensemble_seed6_kd_a05_t1_m1ensemble trung bình nhân trên 1 phương pháp · nhãn m1 –0.067m1 95.76%85.93%94.62%98.98%88.89%75.00%94.66%85.07%95.08%98.97%89.77%50.00%
34 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_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-v20.073m1 95.43% ±0.3482.02% ±0.1194.87% ±0.4498.64% ±1.5689.63% ±1.2880.56% ±12.7394.40% ±0.0980.81% ±0.3795.08% ±0.0098.63% ±1.5790.53% ±1.3161.11% ±25.46
35 e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 unsloth/gemma-3-1b-pt0.064m1 95.07% ±0.5985.68% ±0.5993.33% ±2.9196.94% ±1.0286.30% ±6.7975.00% ±0.0093.82% ±0.7284.72% ±0.6993.44% ±2.8496.91% ±1.0387.12% ±6.9450.00% ±0.00
36 e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-base0.031m1 93.86% ±0.8986.19% ±2.3295.13% ±1.7898.30% ±0.5984.44% ±2.9463.89% ±4.8192.24% ±1.1285.00% ±2.5694.81% ±1.8998.28% ±0.6085.23% ±3.0127.78% ±9.62
37 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.061m1 94.19% ±0.6384.84% ±3.2393.59% ±1.1896.60% ±1.1883.33% ±5.0977.78% ±4.8192.90% ±0.5583.72% ±3.4593.72% ±1.2596.56% ±1.1984.09% ±5.2155.56% ±9.62
38 e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1 Qwen/Qwen3-4B-Base0.131m1 96.20% ±1.1085.81% ±1.2894.36% ±2.7098.98% ±1.0289.63% ±1.7086.11% ±4.8195.29% ±1.3484.86% ±1.4993.99% ±2.8898.97% ±1.0390.53% ±1.7472.22% ±9.62
39 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_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 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.043m1 94.50% ±0.5084.59% ±0.5192.05% ±2.4797.96% ±1.0287.04% ±4.4972.22% ±9.6293.05% ±0.5683.87% ±0.3391.80% ±2.1797.94% ±1.0387.88% ±4.5944.44% ±19.25
40 e2_gemma3_1b_pt_cross_ce_real_lora_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.053m1 94.83% ±0.5385.36% ±1.5093.85% ±0.0095.92% ±1.0284.81% ±2.5775.00% ±0.0093.59% ±0.6184.58% ±1.5794.54% ±0.4795.88% ±1.0385.61% ±2.6250.00% ±0.00
41 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_kd_a05_t1_tsa_m1 vinai/phobert-base-v20.056m1 94.39% ±0.1784.26% ±0.6894.62% ±1.3397.96% ±0.0086.67% ±2.9486.11% ±4.8192.90% ±0.2283.37% ±0.7794.81% ±1.2597.94% ±0.0087.50% ±3.0172.22% ±9.62
42 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-v20.037m1 94.92% ±1.1284.20% ±1.8492.82% ±1.7897.28% ±0.5987.78% ±1.9283.33% ±8.3393.74% ±1.3983.08% ±1.9192.90% ±2.3797.25% ±0.6088.64% ±1.9766.67% ±16.67
43 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_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, nhãn m1 vinai/phobert-base-v20.085m1 95.85% ±0.4584.78% ±2.0191.54% ±0.7798.98% ±0.0087.78% ±1.9280.56% ±12.7394.91% ±0.5283.65% ±2.1891.80% ±0.8298.97% ±0.0088.64% ±1.9761.11% ±25.46
44 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_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, nhãn m1 vinai/phobert-base-v20.028m1 94.74% ±0.9783.24% ±2.0393.08% ±1.3397.62% ±0.5985.56% ±1.1180.56% ±20.9793.39% ±1.2682.30% ±2.2593.17% ±0.9597.59% ±0.6086.36% ±1.1461.11% ±41.94
45 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.082m1 94.88% ±0.6284.84% ±3.1791.79% ±3.8795.92% ±1.0282.59% ±3.3980.56% ±4.8193.51% ±0.8683.72% ±3.2391.80% ±3.7695.88% ±1.0383.33% ±3.4761.11% ±9.62
46 e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1, nhãn m1 jhu-clsp/mmBERT-base0.057m1 94.04% ±0.4084.59% ±1.6792.56% ±2.3595.58% ±2.5783.70% ±3.3988.89% ±4.8192.67% ±0.4083.51% ±2.0992.62% ±2.1795.53% ±2.5984.47% ±3.4777.78% ±9.62
47 e2_qwen3_17b_base_cross_ce_real_lora_m1Qwen3-1.7B Base · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật, nhãn m1 Qwen/Qwen3-1.7B-Base0.101m1 94.35% ±0.7785.04% ±0.6891.54% ±2.7797.96% ±2.0488.89% ±2.2280.56% ±9.6292.99% ±0.9283.87% ±0.7591.80% ±2.9697.94% ±2.0689.77% ±2.2761.11% ±19.25
48 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1 vinai/phobert-base-v20.068m1 95.45% ±0.5684.46% ±2.1693.85% ±1.2998.98% ±0.6589.26% ±1.9576.39% ±9.7494.31% ±0.7683.55% ±2.2494.13% ±1.4198.97% ±0.6590.15% ±1.9952.78% ±19.48
49 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_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, nhãn m1 vinai/phobert-base-v20.052m1 94.61% ±0.8384.78% ±3.9393.33% ±1.1897.96% ±1.0287.78% ±2.2275.00% ±14.4393.28% ±0.9683.80% ±4.1593.17% ±1.7197.94% ±1.0388.26% ±1.7450.00% ±28.87
50 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-v20.058m1 94.61% ±0.1584.65% ±0.7892.31% ±0.7797.28% ±2.1287.41% ±3.3980.56% ±12.7393.22% ±0.1883.72% ±0.9692.62% ±0.0097.25% ±2.1588.26% ±3.4761.11% ±25.46
51 e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.056m1 94.92% ±0.4483.62% ±1.8693.08% ±0.7796.60% ±1.1884.81% ±3.9080.56% ±4.8193.68% ±0.6582.52% ±1.9093.72% ±1.7196.56% ±1.1985.61% ±3.9961.11% ±9.62
52 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-3-270m0.098m1 94.19% ±0.5781.89% ±1.7390.77% ±1.3395.58% ±0.5980.37% ±6.1283.33% ±8.3392.84% ±0.9980.45% ±1.9291.26% ±1.2595.53% ±0.6081.06% ±6.2666.67% ±16.67
53 e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.029m1 95.10% ±0.5786.90% ±1.0293.08% ±1.5498.64% ±0.5984.07% ±2.3172.22% ±9.6293.97% ±0.6986.00% ±1.1793.72% ±1.2598.63% ±0.6084.85% ±2.3744.44% ±19.25
54 e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.034m1 92.87% ±1.6686.00% ±1.5794.10% ±1.1897.28% ±1.1884.07% ±1.7094.44% ±4.8192.24% ±0.9585.00% ±1.6193.99% ±2.0697.25% ±1.1984.85% ±1.7488.89% ±9.62
55 e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.051m1 93.55% ±0.2386.58% ±0.9593.08% ±0.7793.88% ±5.4082.59% ±2.5788.89% ±12.7392.47% ±0.2285.57% ±1.0593.72% ±0.9593.81% ±5.4683.33% ±2.6277.78% ±25.46
56 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.040m1 93.81% ±0.9484.52% ±2.1994.10% ±0.8995.92% ±2.7081.85% ±5.5972.22% ±4.8192.24% ±1.1283.37% ±2.1693.99% ±0.4795.88% ±2.7382.58% ±5.7244.44% ±9.62
57 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.057m1 94.30% ±1.4885.29% ±0.2993.59% ±1.1896.60% ±1.5684.07% ±2.3183.33% ±0.0093.13% ±1.2284.08% ±0.4493.72% ±1.2596.56% ±1.5784.85% ±2.3766.67% ±0.00
58 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-base0.070m1 95.10%86.90%92.31%96.94%83.33%91.67%94.05%85.93%92.62%96.91%84.09%83.33%
59 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1factorized · train thật, nhãn m1 jhu-clsp/mmBERT-base0.047m1 94.39% ±0.7785.61% ±0.8791.28% ±0.8996.26% ±2.5784.07% ±3.2175.00% ±8.3393.02% ±1.0284.58% ±0.8691.53% ±0.9596.22% ±2.5984.85% ±3.2850.00% ±16.67
60 e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1gen_label · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 Qwen/Qwen3-1.7B-Base0.095m1 95.45% ±0.2786.77% ±3.0091.28% ±1.9496.94% ±1.7788.52% ±2.3180.56% ±12.7394.22% ±0.3185.57% ±3.2090.98% ±1.6496.91% ±1.7989.39% ±2.3766.67% ±16.67
61 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-base0.012m1 93.64%83.04%94.62%98.98%87.78%66.67%91.72%82.09%95.08%98.97%88.64%33.33%
62 e3_ensemble_seed_kd_a05_t1_m1ensemble trung bình nhân trên 1 phương pháp · nhãn m1 –0.078m1 95.49%86.13%94.62%98.98%91.11%75.00%94.40%85.50%95.08%98.97%92.05%50.00%
63 e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1gemma-3-1b-pt · 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 unsloth/gemma-3-1b-pt0.042m1 94.19% ±0.8984.71% ±1.1693.85% ±2.0495.92% ±1.7782.22% ±1.9280.56% ±9.6292.99% ±0.9583.65% ±1.2994.26% ±2.1795.88% ±1.7982.95% ±1.9761.11% ±19.25
64 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_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, nhãn m1 vinai/phobert-base-v20.024m1 94.37% ±0.1383.24% ±1.4594.10% ±1.1899.66% ±0.5986.30% ±6.5172.22% ±9.6292.87% ±0.2582.37% ±1.5093.99% ±0.9599.66% ±0.6087.12% ±6.6644.44% ±19.25
65 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_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 2e-05, 6 epoch · train thật, nhãn m1 vinai/phobert-base-v20.031m1 95.34% ±0.6082.79% ±2.6692.05% ±2.2297.96% ±1.7785.19% ±6.4275.00% ±14.4394.17% ±0.7881.73% ±2.7592.35% ±2.3797.94% ±1.7985.98% ±6.5650.00% ±28.87
66 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t2_tsa_m1 vinai/phobert-base-v20.070m1 95.21% ±0.5584.84% ±1.8791.54% ±2.3198.30% ±0.5988.15% ±3.3988.89% ±4.8193.85% ±0.7284.08% ±2.0792.35% ±2.0698.28% ±0.6088.64% ±3.4177.78% ±9.62
67 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_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-v20.031m1 94.94% ±0.9684.71% ±0.5994.10% ±0.4497.62% ±1.1884.44% ±3.3375.00% ±8.3393.62% ±1.2983.87% ±0.5494.26% ±0.8297.59% ±1.1985.23% ±3.4150.00% ±16.67
68 e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 5e-05, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.096m1 94.79% ±1.0085.04% ±2.1290.77% ±0.0096.60% ±1.1887.41% ±3.3955.56% ±26.7993.51% ±1.3384.15% ±2.5790.71% ±0.4796.56% ±1.1987.88% ±3.6538.89% ±9.62
69 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.038m1 94.35% ±0.8583.62% ±2.1994.10% ±1.1897.96% ±1.7784.81% ±3.9080.56% ±4.8192.82% ±1.0482.52% ±2.2695.08% ±0.8297.94% ±1.7985.61% ±3.9961.11% ±9.62
70 e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1factorized · train thật + synth ×1, nhãn m1 jhu-clsp/mmBERT-base0.061m1 92.98% ±0.5085.29% ±1.9893.85% ±2.0496.26% ±3.1284.44% ±3.3380.56% ±20.9791.78% ±0.4384.29% ±2.1594.26% ±2.1796.22% ±3.1585.23% ±3.4161.11% ±41.94
71 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.017m1 95.12% ±0.1985.48% ±1.1892.56% ±1.7897.62% ±0.5986.30% ±2.3158.33% ±0.0093.88% ±0.3184.51% ±1.4293.17% ±1.7197.59% ±0.6087.12% ±2.3716.67% ±0.00
72 e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.026m1 95.56% ±0.6086.19% ±1.6290.77% ±1.3398.30% ±0.5985.93% ±0.6466.67% ±8.3394.48% ±0.6585.29% ±1.6791.80% ±1.4298.28% ±0.6086.74% ±0.6633.33% ±16.67
73 e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_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 MLP 512 · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.065m1 94.88% ±0.6684.20% ±1.5492.56% ±1.1897.28% ±1.5686.30% ±0.6477.78% ±4.8193.62% ±0.9183.23% ±1.6092.62% ±0.8297.25% ±1.5786.74% ±1.3155.56% ±9.62
74 e2_xlmr_base_cross_logitadjust_real_frozenemb_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE điều chỉnh logit theo prior (τ=1) · lr 2e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.018m1 93.66% ±0.6281.18% ±1.5793.59% ±1.1898.98% ±0.0087.41% ±2.5775.00% ±0.0091.87% ±0.7579.96% ±1.4993.99% ±1.8998.97% ±0.0087.88% ±2.3750.00% ±0.00
75 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.026m1 94.94% ±1.0089.08% ±0.7392.05% ±1.6097.28% ±1.5685.93% ±1.2883.33% ±0.0093.62% ±1.3588.27% ±0.7491.80% ±2.1797.25% ±1.5786.74% ±1.3166.67% ±0.00
76 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1factorized · train thật, nhãn m1 jhu-clsp/mmBERT-base0.072m1 95.23% ±0.9186.26% ±1.9389.23% ±2.6695.92% ±2.7081.48% ±6.7075.00% ±8.3393.97% ±1.1085.29% ±2.0689.34% ±2.8495.88% ±2.7382.20% ±6.8550.00% ±16.67
77 e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 3e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.082m1 94.81% ±0.6886.06% ±1.7593.08% ±1.5494.56% ±2.5781.11% ±7.7877.78% ±9.6293.51% ±0.8284.93% ±1.8493.44% ±1.6494.50% ±2.5981.82% ±7.9555.56% ±19.25
78 e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1Qwen3-1.7B Base · 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 Qwen/Qwen3-1.7B-Base0.067m1 93.17%82.85%93.08%97.96%85.56%83.33%91.55%81.66%93.44%97.94%86.36%66.67%
79 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/visobert0.037m1 90.85%82.08%93.08%93.88%84.44%75.00%89.22%80.60%93.44%93.81%85.23%50.00%
80 e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.027m1 92.45%86.51%89.23%98.98%85.56%75.00%90.78%85.29%90.16%98.97%86.36%50.00%
81 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-base0.057m1 94.41% ±0.5485.68% ±3.0793.59% ±1.9496.26% ±0.5983.33% ±3.3366.67% ±8.3393.28% ±0.3884.79% ±3.3593.72% ±1.7196.22% ±0.6084.09% ±3.4133.33% ±16.67
82 e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_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 2e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 vinai/phobert-base-v20.206m1 96.09% ±1.2686.58% ±3.3888.72% ±5.0098.30% ±2.1288.15% ±0.6486.11% ±9.6295.11% ±1.5985.57% ±3.6388.80% ±4.9498.28% ±2.1589.02% ±0.6672.22% ±19.25
83 e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_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 cổng trộn u/v, head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.034m1 95.47% ±0.2785.36% ±0.6990.51% ±1.1899.32% ±0.5988.52% ±1.2875.00% ±14.4394.22% ±0.3884.43% ±0.9890.71% ±1.2599.31% ±0.6088.64% ±1.9750.00% ±28.87
84 e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_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 2e-05, 4 epoch · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1 vinai/phobert-base-v20.052m1 95.91% ±1.1284.26% ±2.1191.54% ±4.2898.64% ±0.5983.33% ±5.0966.67% ±0.0094.80% ±1.4083.23% ±2.1691.53% ±4.6698.63% ±0.6084.09% ±5.2133.33% ±0.00
85 e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_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 + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1 vinai/phobert-base-v20.033m1 95.69% ±0.7484.14% ±2.3992.56% ±2.4797.96% ±0.0082.59% ±5.2566.67% ±8.3394.68% ±1.0183.23% ±2.5492.62% ±2.8497.94% ±0.0083.33% ±5.3733.33% ±16.67
86 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.144m1 94.19% ±0.5586.32% ±0.8491.54% ±0.7793.54% ±2.3685.56% ±1.9275.00% ±8.3392.61% ±0.6485.36% ±0.8991.53% ±1.2593.47% ±2.3886.36% ±1.9750.00% ±16.67
87 e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.082m1 94.23% ±1.5085.10% ±1.9591.03% ±4.6498.98% ±1.0283.70% ±2.8097.22% ±4.8192.90% ±1.8684.01% ±2.0390.98% ±4.5698.97% ±1.0384.47% ±2.8694.44% ±9.62
88 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.108m1 95.54% ±0.8886.64% ±1.1391.79% ±2.3596.26% ±1.5682.59% ±3.5780.56% ±4.8194.37% ±1.0785.71% ±1.1392.08% ±2.3796.22% ±1.5783.71% ±4.3061.11% ±9.62
89 e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-0.6B-Base0.075m1 94.12% ±1.2483.94% ±2.1793.59% ±1.6098.30% ±0.5989.63% ±2.3186.11% ±4.8192.73% ±1.6682.59% ±2.2893.44% ±1.6498.28% ±0.6090.53% ±2.3772.22% ±9.62
90 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t1_tsa_m1 vinai/phobert-base-v20.070m1 94.94% ±0.1784.26% ±1.7493.59% ±1.7898.98% ±1.0285.93% ±3.9086.11% ±9.6293.53% ±0.1783.37% ±1.9293.72% ±1.7198.97% ±1.0386.74% ±3.9972.22% ±19.25
91 e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1Qwen3-0.6B Base · 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 Qwen/Qwen3-0.6B-Base0.077m1 92.51% ±0.9383.30% ±3.6190.26% ±3.8796.26% ±1.1887.41% ±2.8077.78% ±9.6291.61% ±2.0382.02% ±3.8790.44% ±4.1396.22% ±1.1988.26% ±2.8655.56% ±19.25
92 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_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-v20.093m1 94.54% ±1.6087.15% ±3.5892.56% ±0.4499.66% ±0.5985.93% ±1.2888.89% ±12.7393.16% ±1.9386.28% ±3.7392.62% ±0.8299.66% ±0.6086.74% ±1.3177.78% ±25.46
93 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_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, nhãn m1 vinai/phobert-base-v20.104m1 95.52% ±1.1485.93% ±2.0491.54% ±3.0898.98% ±1.0283.33% ±3.3377.78% ±9.6294.37% ±1.4385.07% ±2.0690.98% ±3.2898.97% ±1.0384.09% ±3.4155.56% ±19.25
94 e2_xlmr_base_cross_ce_real_frozenemb_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.044m1 94.01% ±1.0183.30% ±4.0790.26% ±3.5597.28% ±1.1887.78% ±1.1169.44% ±12.7392.47% ±1.2682.16% ±4.7290.98% ±3.5797.25% ±1.1989.02% ±0.6638.89% ±25.46
95 e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE trọng số căn nghịch tần suất · lr 2e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.019m1 94.08% ±0.9682.85% ±1.2092.56% ±3.2098.64% ±0.5986.30% ±2.8083.33% ±8.3392.41% ±1.2181.52% ±1.3793.44% ±2.4698.63% ±0.6087.12% ±2.8666.67% ±16.67
96 e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_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 điều chỉnh logit theo prior (τ=0.25) · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.049m1 94.72% ±0.2883.62% ±1.8493.08% ±1.3397.62% ±1.5686.30% ±1.7077.78% ±9.6293.42% ±0.4082.66% ±2.0593.17% ±1.2597.94% ±1.0387.12% ±1.7455.56% ±19.25
97 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_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 1e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.159m1 95.32% ±0.6884.20% ±2.7891.03% ±1.6098.98% ±1.0284.44% ±5.0961.11% ±31.5594.20% ±1.0483.16% ±2.9391.53% ±1.8998.97% ±1.0385.23% ±5.2150.00% ±16.67
98 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsa_m1 vinai/phobert-base-v20.085m1 95.38% ±0.9084.97% ±0.8492.05% ±2.3599.32% ±0.5987.04% ±3.3980.56% ±12.7394.08% ±1.0884.15% ±0.8692.08% ±2.0699.31% ±0.6087.88% ±3.4761.11% ±25.46
99 e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1 Qwen/Qwen3-Embedding-4B0.061m1 93.94% ±0.5281.02% ±2.3290.00% ±3.2688.78% ±8.6678.89% ±4.7183.33% ±11.7992.24% ±0.6179.64% ±2.2689.34% ±3.4888.66% ±8.7578.98% ±4.0266.67% ±23.57
100 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.025m1 95.83%84.59%89.23%97.96%82.22%16.67%94.66%83.80%90.16%97.94%82.95%16.67%
101 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1Qwen3-Embedding-0.6B · cross-attention (Q = câu KH, K/V = câu bot), backbone đóng băng · CE trọng số căn nghịch tần suất, label smoothing 0.05 · lr 0.0003, 40 epoch · train thật, nhãn m1 Qwen/Qwen3-Embedding-0.6B0.152m1 94.23% ±0.5383.82% ±3.3684.10% ±0.4493.88% ±0.0079.63% ±4.2186.11% ±4.8192.67% ±0.6982.44% ±3.6583.33% ±0.4793.81% ±0.0080.30% ±4.3072.22% ±9.62
102 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_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 5e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.035m1 94.61% ±1.0684.84% ±2.3193.59% ±1.6098.30% ±1.5687.41% ±2.8077.78% ±4.8193.22% ±1.2683.80% ±2.6793.72% ±1.8998.28% ±1.5788.26% ±2.8655.56% ±9.62
103 e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-base0.052m1 93.57%84.97%90.00%96.94%83.33%25.00%91.90%83.80%90.16%96.91%84.09%33.33%
104 e2_visobert_cross_ce_real_frozenemb_lr2e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 uitnlp/visobert0.027m1 92.18%79.96%90.00%94.90%86.67%58.33%90.60%78.25%90.98%94.85%87.50%16.67%
105 e2_visobert_cross_ce_real_frozenemb_lr1e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 uitnlp/visobert0.186m1 92.31%82.08%86.15%94.90%76.67%66.67%90.52%81.02%86.07%94.85%76.14%33.33%
106 e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.161m1 93.77%83.82%90.00%97.96%82.22%58.33%91.98%82.94%90.98%97.94%82.95%16.67%
107 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-0.6B0.149m1 94.37%79.58%84.62%89.80%78.89%66.67%93.19%78.04%84.43%89.69%79.55%33.33%
108 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-IDF0.207m1 95.10%84.78%81.54%94.90%78.89%66.67%93.79%83.80%81.15%94.85%79.55%33.33%
109 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-IDF0.164m1 95.03%84.39%80.77%94.90%78.89%66.67%93.71%83.37%80.33%94.85%79.55%33.33%
110 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-IDF0.280m1 93.24%82.66%84.62%94.90%77.78%83.33%91.72%81.45%84.43%94.85%78.41%66.67%
111 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-IDF0.178m1 94.43%85.36%83.85%94.90%78.89%66.67%93.02%84.22%83.61%94.85%79.55%33.33%
112 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-IDF0.267m1 93.51%81.70%87.69%94.90%76.67%75.00%91.64%80.38%87.70%94.85%77.27%50.00%
113 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-IDF0.298m1 93.37%82.85%84.62%94.90%78.89%83.33%92.07%81.66%84.43%94.85%79.55%66.67%
114 e2_dual_qwen3emb06b_concat_ce_real_frozen_m1Qwen3-Embedding-0.6B · dual encoder (câu bot và câu KH mã hoá riêng, pooling last_token), backbone đóng băng · fusion [u; v; |u−v|; u·v], head MLP 512 · CE · lr head 0.001, 30 epoch · train thật, nhãn m1 Qwen/Qwen3-Embedding-0.6B0.227m1 93.11% ±1.4077.65% ±0.5885.90% ±1.1891.84% ±3.0668.52% ±6.7966.67% ±8.3391.44% ±1.6976.05% ±0.5485.52% ±0.9591.75% ±3.0968.94% ±6.9433.33% ±16.67
115 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-IDF0.282m1 93.31%82.08%83.85%94.90%77.78%83.33%91.72%80.81%83.61%94.85%78.41%66.67%
116 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-IDF0.190m1 94.50%85.36%83.08%94.90%77.78%66.67%93.02%84.22%82.79%94.85%78.41%33.33%
117 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-IDF0.310m1 93.31%83.04%85.38%94.90%80.00%83.33%91.64%81.88%85.25%94.85%80.68%66.67%
118 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-IDF0.356m1 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/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_bamibert_cross_ce_real_frozenemb_lr2e5_m1 ––m1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_bamibert_cross_ce_real_frozenemb_lr4e5_m1 ––m1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e3_fbverifier_devtop1_familygeomean_fbf1_m1fallback_verifier · train thật, nhãn m1 –hardm1 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_m1fallback_verifier · train thật, nhãn m1 –hardm1 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_allensemble trung bình trên 23 phương pháp · nhãn v1 –0.065v1 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_allensemble chọn tham lam Caruana trên 23 phương pháp · nhãn v1 –0.070v1 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_lr4e5XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.023v1 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-v20.011v1 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_frozenembXLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE trọng số căn nghịch tần suất · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.021v1 93.81% ±0.1084.54% ±1.4384.35% ±1.5692.34% ±0.7883.11% ±2.0960.00% ±20.0091.74% ±0.2183.29% ±1.2686.21% ±1.9993.52% ±0.8084.04% ±2.1550.00% ±25.00
6 e3_prior_em_mmbert_bottypeEM điều chỉnh prior trên e2_mmbert_base_cross_ce_real_frozenemb_lr2e5, phạm vi theo loại câu bot (transductive) · nhãn v1 –0.082v1 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_globalEM điều chỉnh prior trên e2_mmbert_base_cross_ce_real_frozenemb_lr2e5, phạm vi toàn bộ (transductive) · nhãn v1 –0.082v1 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_lr2e5mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 jhu-clsp/mmBERT-base0.061v1 94.48% ±0.1682.99% ±2.3585.71% ±1.0289.19% ±1.3583.56% ±4.9486.67% ±11.5592.71% ±0.2981.65% ±2.6188.51% ±1.1590.28% ±1.3984.51% ±5.0883.33% ±14.43
9 e2_xlmr_base_cross_ce_real_frozenembXLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.035v1 94.45% ±0.1285.66% ±1.5082.99% ±3.5891.44% ±0.7881.28% ±2.8560.00% ±0.0092.66% ±0.2484.63% ±1.6684.67% ±4.0492.59% ±0.8082.16% ±2.9350.00% ±0.00
10 e2_xlmr_base_cross_logitadjust_real_frozenembXLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE điều chỉnh logit theo prior (τ=1) · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.035v1 93.74% ±0.6183.85% ±0.9185.37% ±4.1291.89% ±1.3583.11% ±2.0966.67% ±11.5591.60% ±0.8382.61% ±1.0187.36% ±3.0493.06% ±1.3984.04% ±2.1558.33% ±14.43
11 e2_qwen3_06b_base_cross_ce_real_loraQwen3-0.6B Base · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn v1 Qwen/Qwen3-0.6B-Base0.098v1 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_allensemble stacking LR trên 23 phương pháp · nhãn v1 –0.088v1 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_loraQwen3-1.7B Base · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật, nhãn v1 Qwen/Qwen3-1.7B-Base0.083v1 94.01% ±0.2783.51% ±1.2883.67% ±2.0490.09% ±0.7883.11% ±2.8586.67% ±23.0992.20% ±0.2182.13% ±1.5085.82% ±1.7691.20% ±0.8084.04% ±2.9383.33% ±28.87
14 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 vinai/phobert-base-v20.022v1 94.72% ±0.2784.72% ±1.1384.35% ±1.5691.44% ±0.7882.19% ±2.3746.67% ±11.5593.03% ±0.2983.38% ±1.1686.59% ±1.3392.59% ±0.8083.10% ±2.4433.33% ±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-base0.088v1 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_lr4e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 intfloat/multilingual-e5-base0.028v1 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_frozenembXLM-R base · chỉ câu KH · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.076v1 94.08% ±0.7684.46% ±2.0782.99% ±3.1291.89% ±1.3582.65% ±3.1653.33% ±23.0992.06% ±1.0483.19% ±2.3185.44% ±3.5193.06% ±1.3983.57% ±3.2541.67% ±28.87
18 e2_me5_base_cross_ce_real_frozenemb_lr2e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 intfloat/multilingual-e5-base0.056v1 93.67% ±0.5286.79% ±0.2684.35% ±2.1291.89% ±1.3581.74% ±1.5846.67% ±11.5591.55% ±0.6385.88% ±0.5085.82% ±2.3993.06% ±1.3982.63% ±1.6333.33% ±14.43
19 e2_gemma3_270m_cross_ce_real_loragemma-3-270m · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn v1 unsloth/gemma-3-270m0.116v1 93.37% ±1.3183.59% ±0.3085.03% ±2.5789.19% ±0.0078.08% ±2.3766.67% ±11.5591.32% ±1.4981.94% ±0.4488.12% ±1.7690.28% ±0.0078.87% ±2.4458.33% ±14.43
20 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozenQwen3-Embedding-0.6B · cross-attention (Q = câu KH, K/V = câu bot), backbone đóng băng · CE trọng số căn nghịch tần suất, label smoothing 0.05 · lr 0.0003, 40 epoch · train thật, nhãn v1 Qwen/Qwen3-Embedding-0.6B0.114v1 93.37% ±0.2981.52% ±0.9180.95% ±3.8688.29% ±2.0684.02% ±1.5880.00% ±0.0091.41% ±0.6079.73% ±1.0182.38% ±3.6989.35% ±2.1284.98% ±1.6375.00% ±0.00
21 e2_mmbert_base_cross_ce_real_frozenemb_lr4e5mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 jhu-clsp/mmBERT-base0.075v1 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_lr1e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 vinai/phobert-base-v20.026v1 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-base0.123v1 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_lr1e5mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 jhu-clsp/mmBERT-base0.120v1 93.96% ±0.3685.36% ±1.2880.61% ±2.8991.22% ±0.9684.93% ±5.8160.00% ±0.0092.24% ±0.2084.29% ±1.4383.91% ±3.2592.36% ±0.9885.92% ±5.9850.00% ±0.00
25 e2_visobert_cross_ce_real_frozenemb_lr2e5ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 uitnlp/visobert0.093v1 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_lr4e5ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 uitnlp/visobert0.066v1 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-IDF0.123v1 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-IDF0.256v1 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-IDF0.125v1 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_lr1e5ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 uitnlp/visobert0.232v1 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-IDF0.234v1 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/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_mdeberta_cross_ce_real_frozenemb_lr2e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 ––v1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_mdeberta_cross_ce_real_frozenemb_lr2e5_fp32mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 ––v1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_mdeberta_cross_ce_real_frozenemb_lr4e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 microsoft/mdeberta-v3-baseargmaxv1 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_lr1e5ViDeBERTa · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 Fsoft-AIC/videberta-baseargmaxv1 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_lr2e5ViDeBERTa · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 Fsoft-AIC/videberta-baseargmaxv1 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/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_videberta_cross_ce_real_frozenemb_lr4e5ViDeBERTa · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 Fsoft-AIC/videberta-baseargmaxv1 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_lv21mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v2_1 jhu-clsp/mmBERT-base0.076v2_1 94.18% ±0.8586.61% ±1.0878.57% ±2.0487.84% ±0.0082.65% ±4.4066.67% ±30.5592.57% ±0.7685.88% ±1.0481.99% ±1.7688.89% ±0.0083.57% ±4.5358.33% ±38.19
2 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_lv21PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v2_1 vinai/phobert-base-v20.033v2_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

STTModel_NameBackboneThresholdData clean Affirm_F1clean Negate_F1clean Fallback_F1clean Callback_F1clean Relation_F1clean Repeat_F1reply-unseen Affirm_F1reply-unseen Negate_F1reply-unseen Fallback_F1reply-unseen Callback_F1reply-unseen Relation_F1reply-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_m2 –0.051m2 97.55%92.68%71.88%99.48%92.22%78.26%97.05%92.18%71.56%99.48%92.68%54.55%
2 e3_ensemble_devtop1_familygeomean_tempdev_m2 –0.051m2 97.69%92.90%73.37%100.00%92.22%78.26%97.19%92.43%72.90%100.00%92.68%54.55%
3 e3_ensemble_devtop1_familygeomean_temphard_m2 –0.051m2 97.69%92.90%73.37%100.00%92.22%78.26%97.19%92.43%72.90%100.00%92.68%54.55%
4 e3_ensemble_devtop1_familymean_m2 –0.112m2 97.62%92.69%74.10%100.00%92.86%83.33%97.10%92.20%73.65%100.00%93.33%66.67%
5 e3_ensemble_sweep_k2_m2 –0.110m2 97.91%93.94%81.63%97.96%92.12%83.33%97.53%93.59%81.29%97.94%92.59%66.67%
6 e3_ensemble_sweep_k3_m2 –0.051m2 97.73%93.67%76.34%98.48%92.12%72.00%97.29%93.29%75.75%98.46%92.59%46.15%
7 e3_ensemble_sweep_k4_m2 –0.058m2 97.90%93.56%76.34%98.98%92.77%72.00%97.52%93.16%75.75%98.97%93.25%46.15%
8 e3_ensemble_sweep_k5_m2 –0.111m2 98.18%94.66%82.43%98.48%92.77%86.96%97.88%94.40%82.14%98.46%93.25%72.73%
9 e3_ensemble_sweep_k6_m2 –0.087m2 98.11%94.32%80.00%98.48%92.77%86.96%97.74%94.02%79.86%98.46%93.25%72.73%
10 e3_ensemble_sweep_k7_m2 –0.087m2 98.11%94.32%80.00%98.48%92.77%86.96%97.74%94.02%79.86%98.46%93.25%72.73%
11 e3_ensemble_sweep_k8_m2 –0.087m2 98.11%94.32%80.00%98.48%92.77%86.96%97.74%94.02%79.86%98.46%93.25%72.73%
12 e3_ensemble_top_geomean_m2 –0.073m2 97.76%93.37%76.25%98.98%94.05%76.92%97.33%92.96%75.66%98.97%94.55%57.14%
13 e3_ensemble_top_mean_m2 –0.073m2 97.16%92.02%68.91%100.00%94.05%76.92%96.54%91.57%68.64%100.00%94.55%57.14%
14 e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m2 unsloth/gemma-3-1b-pt0.064m2 96.51% ±0.6792.13% ±1.3868.37% ±6.9997.44% ±0.8990.86% ±1.3974.85% ±6.1995.73% ±0.7891.58% ±1.5468.03% ±6.8997.41% ±0.9091.30% ±1.4141.67% ±17.56
15 e2_me5_base_cross_ce_real_frozenemb_lr4e5_m2 intfloat/multilingual-e5-base0.017m2 95.62%92.02%65.22%97.49%87.12%64.29%94.45%91.33%64.00%97.46%87.50%37.50%
16 e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m2 jhu-clsp/mmBERT-base0.155m2 97.35%92.08%73.83%95.38%88.89%76.92%96.75%91.42%73.05%95.34%89.31%57.14%
17 e2_qwen3_17b_base_genlabel_ce_real_lora_m2 Qwen/Qwen3-1.7B-Base0.081m2 97.10%92.64%69.64%97.46%91.02%78.57%96.37%92.15%68.55%97.44%91.46%62.50%
18 e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m2 Qwen/Qwen3-1.7B-Base0.105m2 97.10% ±0.4492.87% ±0.3669.65% ±0.1097.12% ±1.0493.07% ±0.3886.87% ±1.9596.39% ±0.5592.35% ±0.4768.80% ±0.2497.09% ±1.0593.33% ±0.0875.09% ±3.17
19 e3_ensemble_top_greedy_m2 –0.113m2 97.48%93.04%72.46%99.49%93.41%81.48%96.91%92.59%71.92%99.48%93.90%66.67%
20 e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m2 intfloat/multilingual-e5-large-instruct0.017m2 96.95% ±0.2393.13% ±1.1673.53% ±6.6898.11% ±0.6189.59% ±2.3576.32% ±6.3796.25% ±0.3092.70% ±1.1772.97% ±6.7298.09% ±0.6190.02% ±2.3956.71% ±8.36
21 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m2 jhu-clsp/mmBERT-base0.094m2 96.62% ±0.7392.13% ±1.0469.34% ±6.3597.62% ±0.5689.20% ±1.5872.57% ±7.8595.79% ±0.9491.60% ±1.0968.78% ±6.7797.59% ±0.5689.62% ±1.6041.75% ±22.71
22 e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m2 Qwen/Qwen3-4B-Base0.056m2 97.52% ±0.1691.93% ±0.5772.00% ±1.2498.29% ±0.2993.73% ±1.7074.33% ±2.4797.01% ±0.1891.43% ±0.5171.41% ±1.1498.27% ±0.2993.79% ±1.9452.21% ±5.58
23 e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m2 unsloth/gemma-3-1b-pt0.045m2 96.93% ±0.0191.75% ±0.3168.58% ±1.6497.41% ±0.0490.92% ±2.8776.91% ±6.9496.23% ±0.0191.15% ±0.3468.08% ±1.7297.38% ±0.0491.36% ±2.9053.08% ±9.79
24 e2_gemma3_1b_pt_cross_ce_real_lora_r32_m2 unsloth/gemma-3-1b-pt0.045m2 96.57% ±0.5290.46% ±0.4666.62% ±5.3797.40% ±0.0390.24% ±3.0475.59% ±6.7195.83% ±0.7289.66% ±0.5166.04% ±5.5997.37% ±0.0390.86% ±3.2454.27% ±10.90
25 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m2 jhu-clsp/mmBERT-base0.052m2 96.51% ±0.2291.71% ±1.4268.35% ±3.0495.54% ±1.1589.01% ±3.0772.80% ±4.0995.94% ±0.2591.03% ±1.6068.64% ±2.5295.49% ±1.1689.42% ±3.1155.71% ±5.15
26 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m2 vinai/phobert-base-v20.008m2 97.21%91.99%70.62%96.94%89.29%28.57%96.71%91.65%71.97%96.91%89.70%28.57%
27 e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m2 Qwen/Qwen3-4B-Base0.162m2 97.38%90.05%70.32%98.48%92.40%82.76%96.88%89.23%69.51%98.46%92.86%70.59%
28 e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m2 Qwen/Qwen3-8B-Base0.170m2 97.84%93.33%77.89%97.46%91.33%81.48%97.39%93.04%77.19%97.44%91.76%66.67%
29 e3_ensemble_seed6_kd_a05_t1_m2 –0.067m2 97.42%92.48%74.69%97.98%92.86%81.82%96.85%92.09%74.51%97.96%93.33%60.00%
30 e2_me5_base_cross_ce_real_frozenemb_lr2e5_m2 intfloat/multilingual-e5-base0.031m2 96.53% ±0.4392.09% ±1.1870.29% ±4.0896.98% ±0.5089.60% ±1.4367.59% ±3.4295.66% ±0.5491.46% ±1.2069.34% ±4.1996.95% ±0.5190.02% ±1.4530.91% ±9.45
31 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m2 jhu-clsp/mmBERT-base0.061m2 96.44% ±0.3091.66% ±1.3568.47% ±4.8497.77% ±0.7488.84% ±3.0872.86% ±5.4595.73% ±0.3591.06% ±1.5168.21% ±4.5697.75% ±0.7589.25% ±3.1349.05% ±8.61
32 e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m2 Qwen/Qwen3-4B-Base0.131m2 97.90% ±0.4792.31% ±0.4976.20% ±6.6398.14% ±1.0293.28% ±1.2278.65% ±10.7397.47% ±0.5891.81% ±0.6275.60% ±6.7998.12% ±1.0393.76% ±1.2461.99% ±15.60
33 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.043m2 96.80% ±0.2291.08% ±0.5868.77% ±1.1697.96% ±0.5191.38% ±2.6276.67% ±8.8296.04% ±0.2890.62% ±0.4768.21% ±1.2297.94% ±0.5191.83% ±2.6650.00% ±16.67
34 e2_gemma3_1b_pt_cross_ce_real_lora_m2 unsloth/gemma-3-1b-pt0.053m2 96.97% ±0.3991.42% ±1.0569.99% ±2.4896.90% ±0.5190.67% ±1.4573.16% ±4.6396.34% ±0.5290.88% ±1.0070.12% ±2.2096.87% ±0.5191.11% ±1.4647.85% ±6.03
35 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m2 vinai/phobert-base-v20.085m2 97.32% ±0.2791.67% ±1.1374.08% ±1.4697.82% ±0.2890.57% ±1.9279.17% ±8.3496.83% ±0.4091.02% ±1.1873.95% ±1.8397.79% ±0.2991.18% ±1.6958.15% ±19.02
36 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m2 jhu-clsp/mmBERT-base0.082m2 96.92% ±0.2191.36% ±1.4270.36% ±5.1196.24% ±0.3287.83% ±2.3570.70% ±1.2796.17% ±0.3390.69% ±1.4869.79% ±5.2696.21% ±0.3288.23% ±2.3947.62% ±4.12
37 e2_qwen3_17b_base_cross_ce_real_lora_m2 Qwen/Qwen3-1.7B-Base0.101m2 96.68% ±0.3291.54% ±0.4968.72% ±3.2397.78% ±1.5592.33% ±0.6871.49% ±6.2795.96% ±0.3990.91% ±0.5568.42% ±3.2197.76% ±1.5792.79% ±0.6948.45% ±12.25
38 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m2 vinai/phobert-base-v20.068m2 97.25% ±0.3591.30% ±1.2872.53% ±3.7797.98% ±0.4492.11% ±0.9179.58% ±4.5996.65% ±0.4490.80% ±1.3672.43% ±3.7897.96% ±0.4592.66% ±0.7456.24% ±11.63
39 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m2 vinai/phobert-base-v20.052m2 96.97% ±0.2691.46% ±0.9970.38% ±3.1398.29% ±0.3191.26% ±0.9975.75% ±5.9796.29% ±0.2590.88% ±1.0669.93% ±3.3898.27% ±0.3191.68% ±1.0049.16% ±15.69
40 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m2 vinai/phobert-base-v20.058m2 96.93% ±0.1890.97% ±0.1170.92% ±1.9697.60% ±0.8190.09% ±2.5676.25% ±6.8996.23% ±0.2090.49% ±0.2170.61% ±2.5997.58% ±0.8290.51% ±2.6053.95% ±15.73
41 e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m2 unsloth/gemma-3-1b-pt0.056m2 97.02% ±0.3090.77% ±0.8667.84% ±2.8196.92% ±0.4890.58% ±2.6474.72% ±5.0396.37% ±0.3590.20% ±0.8067.54% ±2.0896.89% ±0.4991.03% ±2.6852.92% ±5.24
42 e2_gemma3_270m_cross_ce_real_lora_m2 unsloth/gemma-3-270m0.098m2 96.22% ±0.3989.35% ±0.9665.98% ±2.2395.42% ±0.4987.47% ±3.9468.13% ±1.2795.36% ±0.5888.45% ±1.1665.76% ±2.6595.37% ±0.4987.86% ±4.0045.69% ±5.14
43 e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m2 intfloat/multilingual-e5-large-instruct0.029m2 97.13% ±0.3792.88% ±0.6171.87% ±0.9997.96% ±1.0289.43% ±1.3073.28% ±3.3896.58% ±0.4692.48% ±0.6471.59% ±0.9497.93% ±1.0389.85% ±1.3245.93% ±6.79
44 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m2 jhu-clsp/mmBERT-base0.040m2 96.40% ±0.7091.61% ±1.4966.20% ±4.1796.56% ±1.1387.70% ±4.4967.88% ±5.1895.54% ±0.8590.95% ±1.5365.44% ±3.8296.53% ±1.1488.27% ±4.3139.27% ±5.99
45 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m2 jhu-clsp/mmBERT-base0.057m2 96.74% ±0.7091.79% ±0.1169.34% ±3.9697.92% ±1.0589.69% ±1.2771.17% ±7.6496.14% ±0.5591.08% ±0.1069.35% ±2.9097.90% ±1.0690.12% ±1.2951.43% ±9.90
46 e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m2 jhu-clsp/mmBERT-base0.070m2 96.88%92.95%73.29%96.94%88.34%81.48%96.31%92.38%73.51%96.91%88.75%66.67%
47 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m2 jhu-clsp/mmBERT-base0.047m2 96.53% ±0.3991.94% ±0.3568.75% ±1.8196.24% ±0.8189.32% ±1.5771.48% ±10.6195.77% ±0.5291.37% ±0.3868.46% ±1.9896.20% ±0.8289.74% ±1.5946.12% ±16.90
48 e2_xlmr_base_khonly_ce_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.012m2 96.10%90.37%67.80%95.57%89.94%69.57%95.19%89.75%67.26%95.52%90.36%36.36%
49 e3_ensemble_seed_kd_a05_t1_m2 –0.078m2 97.19%92.28%74.46%98.48%93.49%81.82%96.62%91.99%74.51%98.46%93.98%60.00%
50 e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m2 unsloth/gemma-3-1b-pt0.042m2 96.47% ±0.3791.17% ±0.7866.71% ±1.8197.08% ±0.5588.53% ±1.3881.56% ±3.1895.78% ±0.4090.54% ±0.9066.76% ±1.5297.05% ±0.5588.94% ±1.4161.99% ±8.62
51 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m2 vinai/phobert-base-v20.024m2 96.78% ±0.1590.66% ±0.6667.36% ±0.6698.31% ±0.5890.89% ±3.7478.62% ±7.4296.03% ±0.1390.19% ±0.6266.87% ±0.6098.30% ±0.5891.34% ±3.7952.39% ±17.75
52 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m2 vinai/phobert-base-v20.031m2 97.01% ±0.3290.80% ±1.5968.56% ±3.9998.12% ±0.7989.66% ±3.4175.65% ±5.0596.36% ±0.3990.25% ±1.6467.96% ±3.9198.10% ±0.8090.09% ±3.4648.89% ±15.40
53 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t2_tsa_m2 vinai/phobert-base-v20.070m2 96.95% ±0.2791.37% ±0.8971.11% ±1.9197.80% ±0.5991.43% ±2.2585.48% ±3.5096.21% ±0.3690.87% ±0.9971.09% ±2.2797.78% ±0.5991.66% ±2.2772.11% ±5.16
54 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m2 vinai/phobert-base-v20.031m2 97.01% ±0.5091.46% ±0.5570.93% ±3.9897.29% ±0.3089.86% ±2.2879.29% ±5.7496.34% ±0.7091.00% ±0.5970.49% ±4.2897.26% ±0.3190.29% ±2.3255.56% ±13.88
55 e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m2 unsloth/gemma-3-1b-pt0.096m2 96.95% ±0.4790.55% ±0.8568.98% ±5.1597.43% ±1.0191.19% ±2.0152.35% ±21.1996.30% ±0.6589.97% ±1.0369.33% ±5.7297.40% ±1.0291.63% ±2.0432.70% ±8.85
56 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m2 intfloat/multilingual-e5-large-instruct0.038m2 96.74% ±0.3290.96% ±1.3568.16% ±3.2098.11% ±0.3289.84% ±2.2876.28% ±1.1195.95% ±0.4190.43% ±1.4867.89% ±3.0698.09% ±0.3290.26% ±2.3254.76% ±4.12
57 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.017m2 97.03% ±0.0492.05% ±0.6170.77% ±1.1498.46% ±0.5290.25% ±1.4270.12% ±3.5196.37% ±0.1291.57% ±0.6770.80% ±1.4198.44% ±0.5290.68% ±1.4425.26% ±3.18
58 e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m2 vinai/phobert-base-v20.026m2 97.26% ±0.2492.15% ±0.7973.50% ±3.9198.14% ±1.1790.00% ±1.2472.62% ±5.1996.71% ±0.2591.76% ±0.7073.56% ±3.7298.12% ±1.1890.43% ±1.2638.89% ±14.70
59 e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.065m2 96.93% ±0.1591.01% ±0.7969.81% ±1.9196.95% ±1.0190.98% ±0.6075.73% ±3.3096.27% ±0.2390.45% ±0.7869.54% ±2.5096.92% ±1.0291.43% ±0.6152.61% ±5.74
60 e2_xlmr_base_cross_logitadjust_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.018m2 96.23% ±0.3189.68% ±0.9765.08% ±2.2096.66% ±1.0189.46% ±1.4471.08% ±1.6095.25% ±0.4389.01% ±1.0164.68% ±2.7696.63% ±1.0289.83% ±1.5145.05% ±1.90
61 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m2 intfloat/multilingual-e5-large-instruct0.026m2 97.05% ±0.3893.90% ±0.2874.76% ±4.4697.43% ±0.5491.53% ±0.6774.07% ±0.0096.34% ±0.5193.51% ±0.3274.04% ±4.3297.41% ±0.5491.99% ±0.6853.33% ±0.00
62 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m2 jhu-clsp/mmBERT-base0.072m2 96.84% ±0.4292.24% ±0.6670.97% ±2.0596.23% ±0.3587.65% ±4.3369.39% ±9.1296.11% ±0.5091.67% ±0.6770.40% ±1.8896.19% ±0.3588.04% ±4.4043.32% ±15.43
63 e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m2 jhu-clsp/mmBERT-base0.082m2 96.66% ±0.0792.22% ±0.6770.72% ±0.8896.35% ±1.4288.06% ±3.6368.25% ±7.9095.89% ±0.1791.56% ±0.7570.40% ±0.7596.31% ±1.4488.46% ±3.6843.33% ±14.53
64 e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m2 Qwen/Qwen3-1.7B-Base0.067m2 96.05%90.49%62.73%96.48%91.46%80.00%95.17%89.73%62.32%96.45%91.93%61.54%
65 e2_visobert_cross_ce_real_frozenemb_lr4e5_m2 uitnlp/visobert0.037m2 94.82%89.22%57.00%94.85%89.02%72.00%93.92%88.42%57.22%94.79%89.44%50.00%
66 e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m2 FacebookAI/xlm-roberta-base0.027m2 95.40%91.91%64.02%97.49%89.82%72.00%94.40%91.33%64.26%97.46%90.24%46.15%
67 e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m2 jhu-clsp/mmBERT-base0.057m2 96.52% ±0.3291.58% ±1.3568.45% ±3.9797.92% ±0.0191.27% ±1.2376.05% ±2.2995.89% ±0.2891.05% ±1.4768.54% ±3.0797.90% ±0.0191.74% ±1.2442.52% ±13.09
68 e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.034m2 97.20% ±0.1091.63% ±0.3773.14% ±2.7497.98% ±0.4990.81% ±0.8177.02% ±6.4996.52% ±0.1991.03% ±0.4972.79% ±3.0897.96% ±0.5090.80% ±0.8751.01% ±15.18
69 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m2 jhu-clsp/mmBERT-base0.144m2 96.43% ±0.2492.10% ±0.5868.71% ±2.7495.81% ±0.9889.42% ±1.0069.11% ±2.3895.57% ±0.2891.51% ±0.5867.78% ±2.6795.76% ±0.9990.01% ±1.0342.06% ±8.36
70 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m2 jhu-clsp/mmBERT-base0.108m2 97.11% ±0.1292.45% ±0.8273.97% ±3.4297.08% ±0.7588.00% ±1.6873.55% ±5.6996.47% ±0.1491.94% ±0.8473.49% ±3.4597.05% ±0.7688.40% ±1.7051.47% ±9.43
71 e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m2 Qwen/Qwen3-0.6B-Base0.075m2 96.58% ±0.5891.15% ±1.2568.44% ±5.8897.31% ±0.7592.83% ±1.6073.79% ±2.2395.81% ±0.8090.43% ±1.4467.95% ±6.1897.28% ±0.7693.30% ±1.6355.16% ±3.17
72 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t1_tsa_m2 vinai/phobert-base-v20.070m2 96.83% ±0.2091.21% ±0.8070.81% ±2.1297.81% ±0.2890.34% ±2.1983.80% ±7.0096.03% ±0.2790.65% ±0.9070.44% ±2.3897.79% ±0.2890.78% ±2.2268.18% ±14.45
73 e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m2 Qwen/Qwen3-0.6B-Base0.077m2 95.36% ±0.4089.94% ±1.8062.82% ±3.8797.75% ±1.2092.51% ±1.9478.84% ±8.7594.75% ±0.9889.17% ±1.9664.63% ±5.9997.73% ±1.2192.99% ±1.9756.88% ±18.62
74 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m2 vinai/phobert-base-v20.104m2 97.11% ±0.3691.57% ±0.7273.84% ±5.9497.97% ±0.8789.13% ±2.6978.97% ±1.7896.42% ±0.4791.01% ±0.6373.34% ±5.9997.95% ±0.8889.55% ±2.7356.23% ±5.82
75 e2_xlmr_base_cross_ce_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.044m2 96.25% ±0.3890.42% ±2.0866.87% ±5.9196.46% ±0.5389.63% ±0.2869.09% ±5.2395.36% ±0.4589.86% ±2.3566.72% ±6.4596.42% ±0.5390.04% ±0.2836.57% ±16.67
76 e2_xlmr_base_cross_wcesqrt_real_frozenemb_m2 FacebookAI/xlm-roberta-base0.019m2 96.29% ±0.2890.20% ±0.6968.60% ±4.2096.98% ±0.4989.36% ±0.7174.97% ±5.6595.36% ±0.3489.42% ±0.8368.09% ±4.5696.95% ±0.4989.77% ±0.7154.27% ±10.90
77 e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m2 vinai/phobert-base-v20.049m2 96.87% ±0.2590.97% ±0.8269.60% ±0.5697.45% ±0.0490.04% ±2.1371.67% ±5.1396.19% ±0.2590.45% ±0.9869.37% ±0.8897.43% ±0.0490.46% ±2.1746.83% ±12.22
78 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m2 vinai/phobert-base-v20.159m2 96.86% ±0.2090.81% ±1.3271.95% ±6.3597.97% ±0.8789.40% ±3.1463.43% ±24.5096.20% ±0.3290.33% ±1.4072.06% ±5.7097.95% ±0.8889.82% ±3.1850.51% ±10.78
79 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsa_m2 vinai/phobert-base-v20.085m2 97.03% ±0.4291.75% ±0.4572.72% ±3.5497.98% ±0.0090.85% ±1.9776.05% ±5.6696.30% ±0.5591.25% ±0.4572.44% ±3.7297.96% ±0.0091.29% ±2.0053.47% ±14.82
80 e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m2 Qwen/Qwen3-Embedding-4B0.061m2 95.49% ±0.0288.10% ±1.2165.62% ±1.4893.69% ±5.2284.57% ±0.0671.46% ±14.6494.24% ±0.1187.14% ±1.2564.47% ±1.9993.62% ±5.2884.56% ±0.5850.83% ±12.96
81 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m2 vinai/phobert-base-v20.025m2 97.17%91.46%71.29%97.46%86.59%25.00%96.51%91.08%72.54%97.44%86.96%22.22%
82 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m2 Qwen/Qwen3-Embedding-0.6B0.152m2 95.88% ±0.3588.84% ±1.1067.44% ±3.9994.04% ±0.7484.80% ±1.8969.76% ±4.6494.79% ±0.4887.86% ±1.2266.26% ±4.4293.98% ±0.7485.14% ±1.9249.22% ±6.76
83 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m2 vinai/phobert-base-v20.035m2 96.92% ±0.5091.72% ±1.1970.59% ±4.6898.13% ±0.7890.48% ±1.5076.73% ±1.6496.24% ±0.5791.12% ±1.3569.90% ±4.6098.11% ±0.7890.91% ±1.5353.90% ±3.62
84 e2_me5_base_cross_ce_real_frozenemb_lr1e5_m2 intfloat/multilingual-e5-base0.052m2 96.07%91.36%65.71%95.00%90.24%33.33%95.16%90.72%66.26%94.95%90.68%36.36%
85 e2_visobert_cross_ce_real_frozenemb_lr2e5_m2 uitnlp/visobert0.027m2 95.12%87.30%58.91%95.88%90.36%58.33%94.19%86.36%59.02%95.83%90.80%16.67%
86 e2_visobert_cross_ce_real_frozenemb_lr1e5_m2 uitnlp/visobert0.186m2 94.96%87.77%59.89%95.88%82.50%61.54%93.80%87.08%59.65%95.83%82.58%28.57%
87 e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m2 FacebookAI/xlm-roberta-base0.161m2 95.65%90.11%68.26%95.00%87.27%51.85%94.52%89.57%68.14%94.95%87.65%13.33%
88 e1_tfidf_hybrid_qwen3emb_lr_m2 Qwen/Qwen3-Embedding-0.6B0.149m2 95.64%85.62%62.57%92.15%85.53%69.57%94.74%84.33%62.50%92.06%85.90%36.36%
89 e1_tfidf_svmcal_ce_real_base_m2 TF-IDF0.178m2 96.21%88.50%68.83%94.85%87.50%69.57%95.30%87.46%68.51%94.79%87.90%36.36%
90 e2_dual_qwen3emb06b_concat_ce_real_frozen_m2 Qwen/Qwen3-Embedding-0.6B0.227m2 94.28% ±0.5984.53% ±0.6661.16% ±2.7494.06% ±1.4276.37% ±4.1477.23% ±4.1792.89% ±0.7283.35% ±0.5960.43% ±2.7794.00% ±1.4476.55% ±4.2244.34% ±15.71
– e3_fbverifier_devtop1_familygeomean_fbf1_m2 –hardm2 98.39%95.26%87.20%97.98%92.86%75.86%98.16%95.17%87.66%97.96%93.33%58.82%
– e3_fbverifier_devtop1_familygeomean_macrof1_m2 –hardm2 98.39%95.26%87.20%97.98%92.86%75.86%98.16%95.17%87.66%97.96%93.33%58.82%
1 e3_ensemble_devtop1_familygeomean_m1ensemble trung bình nhân theo nhóm trên 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051m1 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 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051m1 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 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051m1 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_m1ensemble trung bình đều theo nhóm trên 8 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.112m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 2 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.110m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 3 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.051m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 4 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.058m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.111m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.087m1 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 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.087m1 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_m1ensemble quét tổ hợp nhóm, chọn tổ hợp tốt nhất trên dev trên 5 phương pháp · giữ 1 phương pháp tốt nhất mỗi nhóm theo dev · nhãn m1 –0.087m1 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_m1ensemble trung bình nhân trên 5 phương pháp · nhãn m1 –0.073m1 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_m1ensemble trung bình trên 5 phương pháp · nhãn m1 –0.073m1 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_m1gemma-3-1b-pt · 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 unsloth/gemma-3-1b-pt0.064m1 96.05% ±0.6990.23% ±1.2866.38% ±6.7894.87% ±1.1989.07% ±2.0970.56% ±4.1995.14% ±0.8189.49% ±1.4365.96% ±6.6894.81% ±1.2089.47% ±2.1236.11% ±12.73
15 e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-base0.017m1 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_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.155m1 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_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-1.7B-Base0.081m1 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_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-1.7B-Base0.105m1 96.53% ±0.4990.73% ±0.2366.36% ±0.4695.23% ±1.3691.11% ±0.4486.87% ±1.9595.67% ±0.6089.97% ±0.4865.58% ±0.7295.02% ±1.2791.32% ±0.6875.09% ±3.17
19 e3_ensemble_top_greedy_m1ensemble chọn tham lam Caruana trên 5 phương pháp · nhãn m1 –0.113m1 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_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 3e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.017m1 96.49% ±0.2390.67% ±1.3070.55% ±5.8895.03% ±0.5089.06% ±1.3175.30% ±5.6995.66% ±0.3189.97% ±1.3369.86% ±5.8894.98% ±0.5089.46% ±1.3355.25% ±7.42
21 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m1 jhu-clsp/mmBERT-base0.094m1 96.21% ±0.7189.77% ±1.0566.22% ±6.1895.66% ±0.7288.45% ±1.3972.57% ±7.8595.26% ±0.9189.00% ±1.1165.54% ±6.5495.61% ±0.7288.84% ±1.4241.75% ±22.71
22 e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1factorized · train thật + synth ×1, nhãn m1 jhu-clsp/mmBERT-base0.108m1 95.27% ±0.6288.97% ±1.3761.31% ±4.4893.09% ±0.4686.38% ±2.0477.59% ±4.4894.51% ±0.6288.09% ±1.5561.77% ±3.9793.02% ±0.4786.73% ±2.0656.86% ±11.89
23 e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-4B-Base0.056m1 97.03% ±0.1689.56% ±0.5769.10% ±1.3896.51% ±0.4792.31% ±0.6871.58% ±2.4396.38% ±0.1788.81% ±0.4768.33% ±1.3796.47% ±0.4792.33% ±0.9148.73% ±5.35
24 e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 5 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.045m1 96.46% ±0.0389.55% ±0.0865.70% ±1.1295.19% ±0.2789.60% ±2.8676.91% ±6.9495.63% ±0.0288.68% ±0.1665.19% ±1.5695.14% ±0.2790.00% ±2.8953.08% ±9.79
25 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-v20.033m1 96.09% ±0.1788.59% ±0.7363.41% ±1.6295.25% ±1.0788.42% ±1.9369.53% ±7.5995.33% ±0.3087.79% ±0.8463.21% ±1.8195.20% ±1.0888.97% ±2.0936.06% ±16.27
26 e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=32 · CE · lr 0.0001, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.045m1 96.10% ±0.5088.15% ±0.4564.09% ±5.2795.42% ±0.8289.34% ±3.0374.66% ±7.5995.20% ±0.7387.24% ±0.5463.26% ±5.5495.54% ±0.7589.90% ±3.2053.17% ±12.22
27 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 6 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.052m1 96.09% ±0.2189.27% ±1.7065.49% ±2.8293.01% ±1.0288.05% ±2.7071.41% ±4.7795.40% ±0.2188.34% ±1.9165.57% ±2.3792.94% ±1.0388.43% ±2.7453.90% ±3.62
28 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-base0.051m1 96.55% ±0.3290.28% ±1.1867.29% ±2.5895.11% ±0.7688.75% ±0.6077.49% ±3.7495.80% ±0.3089.51% ±1.2666.64% ±2.3795.06% ±0.7689.14% ±0.6059.05% ±6.87
29 e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 jhu-clsp/mmBERT-base0.118m1 96.64% ±0.3088.85% ±1.2366.47% ±4.3294.79% ±1.1588.49% ±1.2476.05% ±8.1695.87% ±0.4187.92% ±1.3065.77% ±4.6494.74% ±1.1688.87% ±1.2556.07% ±14.15
30 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1PhoBERT-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 m1 vinai/phobert-base-v20.008m1 96.60%89.27%67.60%94.06%87.50%28.57%95.86%88.64%68.86%94.00%87.86%28.57%
31 e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1 Qwen/Qwen3-4B-Base0.162m1 96.97%87.62%67.02%96.55%90.71%80.00%96.34%86.54%66.10%96.52%91.11%66.67%
32 e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1 Qwen/Qwen3-8B-Base0.170m1 97.39%91.04%74.07%96.04%90.71%78.57%96.81%90.50%73.20%96.00%91.11%62.50%
33 e3_ensemble_seed6_kd_a05_t1_m1ensemble trung bình nhân trên 1 phương pháp · nhãn m1 –0.067m1 97.01%90.01%71.51%94.63%91.43%81.82%96.32%89.36%71.17%94.58%91.86%60.00%
34 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_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-v20.073m1 96.73% ±0.1288.10% ±0.1568.03% ±0.9595.55% ±0.4989.03% ±3.0074.22% ±1.5796.07% ±0.0587.16% ±0.3567.84% ±0.8795.51% ±0.5089.40% ±3.0650.72% ±9.68
35 e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 unsloth/gemma-3-1b-pt0.064m1 96.65% ±0.0389.29% ±0.5368.75% ±0.3895.48% ±0.8389.18% ±3.3871.15% ±3.3395.85% ±0.0488.56% ±0.5768.20% ±0.3595.44% ±0.8489.57% ±3.4245.24% ±4.12
36 e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-base0.031m1 96.12% ±0.4489.55% ±1.1467.28% ±3.9493.83% ±0.5487.84% ±1.7366.67% ±5.0295.14% ±0.5788.61% ±1.2166.18% ±4.0893.77% ±0.5588.22% ±1.7530.30% ±10.50
37 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.061m1 96.05% ±0.2889.15% ±1.3265.34% ±4.2695.64% ±0.9587.49% ±2.7172.86% ±5.4595.23% ±0.3188.28% ±1.4864.92% ±4.0095.59% ±0.9687.87% ±2.7549.05% ±8.61
38 e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1 Qwen/Qwen3-4B-Base0.131m1 97.35% ±0.5289.96% ±0.5673.13% ±5.8395.91% ±1.4791.50% ±1.4775.71% ±9.8396.76% ±0.6589.27% ±0.6572.46% ±5.8395.87% ±1.4891.93% ±1.4957.58% ±13.12
39 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_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 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.043m1 96.26% ±0.2088.63% ±0.4665.97% ±1.4795.37% ±0.5189.66% ±2.3173.58% ±6.6895.35% ±0.2587.96% ±0.2865.34% ±1.6295.32% ±0.5290.06% ±2.3445.56% ±10.72
40 e2_gemma3_1b_pt_cross_ce_real_lora_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.053m1 96.47% ±0.4189.28% ±0.9867.46% ±2.3595.11% ±0.2488.41% ±1.2973.16% ±4.6395.70% ±0.5188.60% ±0.9567.56% ±2.0895.06% ±0.2488.79% ±1.3147.85% ±6.03
41 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_kd_a05_t1_tsa_m1 vinai/phobert-base-v20.056m1 96.28% ±0.2588.71% ±0.6666.79% ±1.1095.21% ±0.5489.31% ±2.7477.60% ±2.1595.37% ±0.2887.96% ±0.7066.29% ±1.2595.16% ±0.5589.70% ±2.7959.17% ±2.22
42 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-v20.037m1 96.48% ±0.5588.46% ±0.3867.56% ±2.2694.24% ±1.3590.28% ±1.0882.40% ±2.9795.71% ±0.7287.59% ±0.3767.15% ±2.5594.19% ±1.3790.69% ±1.0964.96% ±2.96
43 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_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, nhãn m1 vinai/phobert-base-v20.085m1 96.82% ±0.2789.24% ±1.2471.14% ±1.3694.48% ±0.2789.12% ±1.5277.16% ±6.9996.19% ±0.4088.35% ±1.3270.85% ±1.7094.43% ±0.2789.67% ±1.2955.18% ±16.43
44 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_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, nhãn m1 vinai/phobert-base-v20.028m1 96.49% ±0.4288.07% ±0.5765.24% ±1.2395.83% ±0.7389.02% ±1.0068.48% ±13.4795.68% ±0.5387.32% ±0.7064.59% ±1.0395.79% ±0.7389.41% ±1.0243.47% ±27.61
45 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.082m1 96.44% ±0.2188.85% ±1.4867.30% ±4.6394.32% ±0.2286.42% ±2.3368.29% ±1.4195.55% ±0.3387.92% ±1.5466.61% ±4.7494.26% ±0.2386.77% ±2.3744.79% ±2.12
46 e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật + synth ×1, nhãn m1 jhu-clsp/mmBERT-base0.057m1 96.14% ±0.1889.13% ±1.1463.28% ±1.5094.77% ±0.4187.75% ±0.8872.92% ±5.0895.32% ±0.1988.27% ±1.3962.90% ±1.7594.71% ±0.4188.13% ±0.8854.17% ±7.22
47 e2_qwen3_17b_base_cross_ce_real_lora_m1Qwen3-1.7B Base · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật, nhãn m1 Qwen/Qwen3-1.7B-Base0.101m1 96.15% ±0.2389.31% ±0.4165.82% ±2.8395.68% ±2.0190.73% ±0.9971.49% ±6.2795.27% ±0.2688.46% ±0.4665.38% ±2.8095.64% ±2.0391.14% ±1.0148.45% ±12.25
48 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1 vinai/phobert-base-v20.068m1 96.84% ±0.3388.90% ±1.3169.62% ±3.7994.87% ±0.6590.94% ±1.0579.58% ±4.5996.13% ±0.4388.16% ±1.3869.35% ±3.8094.82% ±0.6591.45% ±0.9056.24% ±11.63
49 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_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, nhãn m1 vinai/phobert-base-v20.052m1 96.43% ±0.2888.75% ±1.0466.97% ±3.1495.52% ±0.5089.95% ±0.7474.78% ±4.6795.60% ±0.2887.89% ±1.1266.35% ±3.3595.48% ±0.5090.31% ±0.8047.77% ±13.38
50 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-v20.058m1 96.36% ±0.2288.34% ±0.0968.04% ±1.9495.17% ±0.7788.88% ±2.4069.89% ±6.5495.50% ±0.2687.58% ±0.2367.59% ±2.5395.12% ±0.7789.26% ±2.4445.97% ±13.64
51 e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.056m1 96.57% ±0.2888.51% ±0.7265.51% ±2.9394.04% ±0.4788.90% ±2.5274.72% ±5.0395.80% ±0.3687.75% ±0.7165.18% ±2.4793.98% ±0.4889.30% ±2.5652.92% ±5.24
52 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-3-270m0.098m1 95.78% ±0.3387.11% ±0.7363.21% ±1.8793.20% ±0.2986.20% ±3.8166.67% ±0.0094.81% ±0.5586.04% ±0.7862.99% ±2.5093.45% ±0.0486.56% ±3.8744.21% ±7.29
53 e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.029m1 96.65% ±0.3490.38% ±0.7268.69% ±0.8096.03% ±0.9888.66% ±1.3271.24% ±1.5995.98% ±0.4189.73% ±0.7768.34% ±0.9595.99% ±0.9889.06% ±1.3443.33% ±5.77
54 e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.034m1 95.59% ±0.9689.56% ±0.8863.43% ±2.2294.72% ±1.3988.15% ±0.6779.39% ±7.5995.17% ±0.6488.78% ±0.9165.03% ±0.4994.67% ±1.4088.54% ±0.6864.77% ±11.30
55 e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.051m1 95.77% ±0.1489.93% ±1.0365.00% ±0.6694.31% ±2.4586.76% ±1.9771.96% ±15.2295.07% ±0.1589.19% ±1.1665.59% ±0.5794.25% ±2.4887.12% ±2.0054.89% ±20.59
56 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.040m1 95.98% ±0.6889.09% ±1.3563.36% ±3.8794.79% ±1.0986.62% ±4.3466.91% ±4.9795.00% ±0.8188.19% ±1.3762.50% ±3.5694.73% ±1.1087.14% ±4.1538.26% ±6.91
57 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.057m1 96.24% ±0.7789.16% ±0.1166.15% ±3.6695.78% ±1.2988.32% ±1.6471.17% ±7.6495.49% ±0.6488.18% ±0.1165.98% ±2.7195.74% ±1.3088.70% ±1.6751.43% ±9.90
58 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-base0.070m1 96.50%90.47%70.18%93.60%87.72%75.86%95.83%89.66%70.19%93.53%88.10%58.82%
59 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1factorized · train thật, nhãn m1 jhu-clsp/mmBERT-base0.047m1 96.15% ±0.3989.43% ±0.4865.36% ±1.6694.49% ±1.0288.31% ±1.7371.48% ±10.6195.27% ±0.5288.61% ±0.5264.90% ±1.6994.43% ±1.0388.69% ±1.7646.12% ±16.90
60 e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1gen_label · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 Qwen/Qwen3-1.7B-Base0.095m1 96.91% ±0.1290.26% ±1.4467.84% ±2.5896.61% ±0.8091.04% ±0.6981.38% ±4.5496.10% ±0.1889.36% ±1.6066.84% ±2.8596.57% ±0.8191.47% ±0.7066.03% ±5.74
61 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-base0.012m1 95.63%88.14%65.95%92.38%89.77%66.67%94.49%87.30%65.17%92.31%90.17%30.77%
62 e3_ensemble_seed_kd_a05_t1_m1ensemble trung bình nhân trên 1 phương pháp · nhãn m1 –0.078m1 96.81%90.03%71.72%95.57%92.66%78.26%96.14%89.51%71.60%95.52%93.10%54.55%
63 e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1gemma-3-1b-pt · 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 unsloth/gemma-3-1b-pt0.042m1 96.02% ±0.4189.30% ±0.7264.84% ±2.1294.32% ±0.4586.72% ±1.5179.29% ±4.8995.22% ±0.4388.53% ±0.8664.89% ±1.7594.26% ±0.4687.07% ±1.5458.66% ±11.29
64 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_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, nhãn m1 vinai/phobert-base-v20.024m1 96.28% ±0.1688.16% ±0.8164.73% ±0.6795.91% ±0.5289.36% ±3.9276.31% ±3.6495.37% ±0.1587.43% ±0.7964.07% ±0.5195.87% ±0.5389.76% ±3.9748.66% ±11.37
65 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_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 2e-05, 6 epoch · train thật, nhãn m1 vinai/phobert-base-v20.031m1 96.52% ±0.3988.27% ±1.4866.12% ±3.7395.21% ±0.7488.39% ±3.3275.65% ±5.0595.76% ±0.5287.50% ±1.5465.50% ±3.7695.16% ±0.7488.77% ±3.3748.89% ±15.40
66 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t2_tsa_m1 vinai/phobert-base-v20.070m1 96.55% ±0.2989.13% ±1.0868.39% ±2.1094.91% ±0.5790.65% ±2.6783.30% ±4.7195.71% ±0.3888.41% ±1.2068.23% ±2.4094.86% ±0.5790.86% ±2.7168.70% ±7.42
67 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_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-v20.031m1 96.52% ±0.4388.94% ±0.4968.09% ±3.6194.25% ±0.2488.18% ±2.3277.00% ±3.8095.71% ±0.6188.22% ±0.5267.52% ±3.8794.20% ±0.2588.56% ±2.3652.03% ±10.99
68 e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1gemma-3-1b-pt · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 5e-05, 3 epoch · train thật, nhãn m1 unsloth/gemma-3-1b-pt0.096m1 96.45% ±0.4588.64% ±1.0466.83% ±4.8295.31% ±1.5789.71% ±2.1049.97% ±21.2295.59% ±0.6187.92% ±1.2566.82% ±5.2795.26% ±1.5889.90% ±2.3230.00% ±8.66
69 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.038m1 96.21% ±0.3488.53% ±1.2265.09% ±2.6995.68% ±0.3587.88% ±2.3075.33% ±1.4595.28% ±0.4387.74% ±1.3264.70% ±2.5395.63% ±0.3588.25% ±2.3353.49% ±3.57
70 e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1factorized · train thật + synth ×1, nhãn m1 jhu-clsp/mmBERT-base0.061m1 95.50% ±0.0389.36% ±1.0362.91% ±1.8494.50% ±1.2788.18% ±1.8471.85% ±10.2594.75% ±0.1788.53% ±1.1163.47% ±2.5794.44% ±1.2888.74% ±1.7047.94% ±24.66
71 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.017m1 96.51% ±0.0789.48% ±0.5967.81% ±1.0895.83% ±0.2688.59% ±1.5070.12% ±3.5195.69% ±0.1288.73% ±0.6767.54% ±1.3995.79% ±0.2788.97% ±1.5325.26% ±3.18
72 e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.026m1 96.76% ±0.2189.49% ±0.6170.29% ±4.0295.38% ±1.0188.39% ±1.2072.62% ±5.1996.07% ±0.2088.82% ±0.5270.24% ±3.8095.34% ±1.0288.77% ±1.2238.89% ±14.70
73 e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_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 MLP 512 · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.065m1 96.42% ±0.1588.58% ±0.7367.34% ±2.0094.24% ±1.0089.27% ±0.8473.00% ±5.8795.61% ±0.2287.78% ±0.7166.89% ±2.6794.18% ±1.0189.45% ±1.1849.29% ±8.07
74 e2_xlmr_base_cross_logitadjust_real_frozenemb_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE điều chỉnh logit theo prior (τ=1) · lr 2e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.018m1 95.80% ±0.3287.17% ±0.7862.71% ±2.2194.19% ±1.2288.24% ±1.5069.23% ±0.0094.64% ±0.4086.24% ±0.6962.10% ±2.6694.13% ±1.2388.56% ±1.5942.86% ±0.00
75 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1multilingual-e5-large-instruct · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-large-instruct0.026m1 96.55% ±0.4891.43% ±0.3471.17% ±4.5595.02% ±0.4390.09% ±0.6573.19% ±1.5395.69% ±0.6590.79% ±0.3670.30% ±4.4094.97% ±0.4390.51% ±0.6652.22% ±1.92
76 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1factorized · train thật, nhãn m1 jhu-clsp/mmBERT-base0.072m1 96.44% ±0.4389.77% ±0.8867.38% ±1.9294.63% ±0.2786.52% ±4.3869.39% ±9.1295.60% ±0.5188.95% ±0.9266.66% ±1.7194.58% ±0.2786.88% ±4.4543.32% ±15.43
77 e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 3e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.082m1 96.27% ±0.0589.90% ±0.6867.56% ±1.1494.71% ±0.8886.43% ±4.0468.25% ±7.9095.38% ±0.1589.01% ±0.7667.09% ±1.1994.65% ±0.8986.78% ±4.1043.33% ±14.53
78 e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1Qwen3-1.7B Base · 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 Qwen/Qwen3-1.7B-Base0.067m1 95.55%88.30%60.65%95.52%89.53%74.07%94.53%87.44%60.32%95.48%89.94%53.33%
79 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/visobert0.037m1 94.52%87.03%55.13%92.46%87.86%72.00%93.54%86.01%55.21%92.39%88.24%50.00%
80 e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.027m1 95.03%89.62%61.87%95.10%88.51%72.00%93.98%88.69%61.97%95.05%88.89%46.15%
81 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-base0.057m1 96.10% ±0.3189.48% ±1.6465.39% ±3.8796.43% ±0.4789.98% ±1.5673.73% ±3.2895.34% ±0.2688.74% ±1.8965.23% ±3.3096.39% ±0.4890.41% ±1.5839.81% ±13.13
82 e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_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 2e-05, 4 epoch · train thật + synth ×1 (synth mù test, lọc theo dev), nhãn m1 vinai/phobert-base-v20.206m1 96.73% ±0.4489.84% ±1.8075.09% ±7.7694.91% ±0.7289.82% ±1.0775.33% ±10.9895.99% ±0.5889.06% ±1.9574.59% ±8.1994.86% ±0.7390.22% ±1.0956.59% ±17.96
83 e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_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 cổng trộn u/v, head linear · CE · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.034m1 96.68% ±0.1289.19% ±0.4170.25% ±2.5095.27% ±0.5589.35% ±0.5273.81% ±6.3095.85% ±0.2188.36% ±0.5269.76% ±2.7995.22% ±0.5689.31% ±0.7246.76% ±14.72
84 e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_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 2e-05, 4 epoch · train thật + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1 vinai/phobert-base-v20.052m1 96.77% ±0.4888.91% ±0.7969.35% ±4.2295.87% ±1.1287.50% ±3.3169.00% ±6.4695.98% ±0.5788.07% ±0.8468.59% ±4.0395.83% ±1.1387.87% ±3.3736.19% ±6.60
85 e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_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 + synth ×1 (khớp tỉ lệ fallback của train thật), nhãn m1 vinai/phobert-base-v20.033m1 96.62% ±0.2888.80% ±1.0669.16% ±2.9695.06% ±1.2487.22% ±3.1170.69% ±8.5295.88% ±0.3688.06% ±1.1668.89% ±2.7495.01% ±1.2587.59% ±3.1537.54% ±17.00
86 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.144m1 96.01% ±0.2389.78% ±0.4865.85% ±2.9294.02% ±0.4988.33% ±1.0069.11% ±2.3895.03% ±0.2888.96% ±0.4764.84% ±2.8393.96% ±0.5088.88% ±1.1342.06% ±8.36
87 e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.082m1 96.12% ±0.7189.37% ±0.6965.76% ±4.4494.95% ±0.9687.58% ±1.6376.08% ±11.1395.28% ±0.8988.56% ±0.7165.23% ±4.3494.90% ±0.9787.95% ±1.6661.52% ±16.17
88 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 jhu-clsp/mmBERT-base0.108m1 96.66% ±0.1489.96% ±0.9770.83% ±3.2995.30% ±1.2086.92% ±1.5072.53% ±3.9695.88% ±0.1789.20% ±0.9970.28% ±3.1995.25% ±1.2187.49% ±1.8850.00% ±7.14
89 e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1gen_label · train thật, nhãn m1 Qwen/Qwen3-0.6B-Base0.075m1 96.13% ±0.5488.87% ±1.2665.39% ±5.9494.61% ±0.9392.36% ±2.0272.95% ±1.3695.21% ±0.7587.89% ±1.3864.76% ±6.2294.55% ±0.9492.81% ±2.0554.07% ±1.28
90 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t1_tsa_m1 vinai/phobert-base-v20.070m1 96.44% ±0.1788.85% ±0.7568.03% ±1.9494.94% ±0.2389.38% ±2.3181.61% ±5.5495.54% ±0.2288.06% ±0.8467.53% ±2.1894.89% ±0.2389.78% ±2.3464.66% ±11.34
91 e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1Qwen3-0.6B Base · 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 Qwen/Qwen3-0.6B-Base0.077m1 94.82% ±0.3287.70% ±1.9660.26% ±3.5995.61% ±1.0690.93% ±1.5875.64% ±8.4194.08% ±0.8986.70% ±2.1361.81% ±5.5895.56% ±1.0791.36% ±1.6152.38% ±17.17
92 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_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-v20.093m1 96.44% ±0.6790.12% ±1.0669.63% ±1.0194.99% ±1.4688.56% ±1.4077.28% ±1.9895.61% ±0.8689.44% ±1.1268.91% ±0.9794.94% ±1.4788.94% ±1.4359.02% ±4.07
93 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_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, nhãn m1 vinai/phobert-base-v20.104m1 96.65% ±0.3589.34% ±0.9171.31% ±6.0095.10% ±1.2887.70% ±2.2778.02% ±3.4295.82% ±0.4588.56% ±0.8570.68% ±6.0595.05% ±1.2988.07% ±2.3154.96% ±5.94
94 e2_xlmr_base_cross_ce_real_frozenemb_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.044m1 95.85% ±0.3487.99% ±2.2464.36% ±5.3693.16% ±0.5388.77% ±0.3769.09% ±5.2394.88% ±0.4187.13% ±2.6864.09% ±5.8593.09% ±0.5389.36% ±0.7336.57% ±16.67
95 e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE trọng số căn nghịch tần suất · lr 2e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.019m1 95.84% ±0.2687.81% ±0.6566.15% ±3.6894.00% ±0.2588.08% ±0.7074.00% ±3.9794.77% ±0.3586.79% ±0.7765.64% ±4.1093.95% ±0.2588.45% ±0.7052.88% ±8.55
96 e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_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 điều chỉnh logit theo prior (τ=0.25) · lr head 0.0001, backbone 2e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.049m1 96.35% ±0.1988.38% ±0.8566.55% ±0.3895.19% ±0.2188.76% ±1.4768.25% ±2.7595.52% ±0.1887.60% ±1.0166.22% ±0.6795.32% ±0.5289.14% ±1.4942.59% ±8.49
97 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_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 1e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.159m1 96.40% ±0.2088.30% ±1.2869.27% ±6.1895.25% ±1.0288.15% ±2.1761.51% ±23.3395.61% ±0.3487.56% ±1.3469.21% ±5.5795.21% ±1.0388.53% ±2.2048.13% ±11.70
98 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsa_m1 vinai/phobert-base-v20.085m1 96.60% ±0.4389.39% ±0.5269.74% ±3.4895.27% ±0.3089.33% ±2.1675.02% ±4.8095.73% ±0.5388.65% ±0.5469.22% ±3.5395.22% ±0.3089.73% ±2.1952.00% ±13.81
99 e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1 Qwen/Qwen3-Embedding-4B0.061m1 95.09% ±0.0986.16% ±1.2162.48% ±1.6893.19% ±5.2184.26% ±0.4471.46% ±14.6493.74% ±0.1985.02% ±1.2561.23% ±2.1493.11% ±5.2784.24% ±0.0550.83% ±12.96
100 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.025m1 96.66%88.78%68.04%94.12%85.06%25.00%95.85%88.12%68.97%94.06%85.38%22.22%
101 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1Qwen3-Embedding-0.6B · cross-attention (Q = câu KH, K/V = câu bot), backbone đóng băng · CE trọng số căn nghịch tần suất, label smoothing 0.05 · lr 0.0003, 40 epoch · train thật, nhãn m1 Qwen/Qwen3-Embedding-0.6B0.152m1 95.43% ±0.3587.18% ±1.4164.46% ±3.7492.62% ±0.7184.44% ±1.6769.76% ±4.6494.22% ±0.4786.06% ±1.5763.18% ±4.1392.55% ±0.7284.76% ±1.6949.22% ±6.76
102 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_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 5e-05, 4 epoch · train thật, nhãn m1 vinai/phobert-base-v20.035m1 96.35% ±0.5189.07% ±1.1868.00% ±4.1094.91% ±1.0389.06% ±1.0673.81% ±2.0695.51% ±0.5988.23% ±1.3867.29% ±4.1094.86% ±1.0489.45% ±1.0850.00% ±0.00
103 e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 intfloat/multilingual-e5-base0.052m1 95.53%88.91%62.57%91.79%88.76%33.33%94.42%88.02%62.68%91.71%89.16%36.36%
104 e2_visobert_cross_ce_real_frozenemb_lr2e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn m1 uitnlp/visobert0.027m1 94.79%85.30%57.07%93.94%88.64%58.33%93.76%84.17%57.07%93.88%89.02%16.67%
105 e2_visobert_cross_ce_real_frozenemb_lr1e5_m1ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 uitnlp/visobert0.186m1 94.60%85.71%58.33%93.47%82.14%61.54%93.29%84.82%57.85%93.40%82.21%28.57%
106 e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn m1 FacebookAI/xlm-roberta-base0.161m1 95.25%88.15%66.48%92.31%86.55%51.85%94.01%87.42%66.27%92.23%86.90%13.33%
107 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-0.6B0.149m1 95.31%84.20%60.27%91.19%85.54%69.57%94.33%82.81%60.06%91.10%85.89%36.36%
108 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-IDF0.207m1 95.57%86.96%67.52%93.00%86.06%72.73%94.49%85.90%66.89%92.93%86.42%40.00%
109 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-IDF0.164m1 95.63%86.90%65.42%93.00%86.06%69.57%94.56%85.84%64.69%92.93%86.42%36.36%
110 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-IDF0.280m1 94.94%85.97%60.27%92.54%85.89%80.00%93.83%84.79%60.06%92.46%86.25%61.54%
111 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-IDF0.178m1 95.70%86.78%66.26%93.00%86.06%69.57%94.65%85.59%65.81%92.93%86.42%36.36%
112 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-IDF0.267m1 95.08%85.74%61.13%92.54%85.19%78.26%93.78%84.53%60.11%92.46%85.53%54.55%
113 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-IDF0.298m1 94.95%86.17%61.45%92.54%86.06%76.92%93.93%85.02%61.68%92.46%86.42%57.14%
114 e2_dual_qwen3emb06b_concat_ce_real_frozen_m1Qwen3-Embedding-0.6B · dual encoder (câu bot và câu KH mã hoá riêng, pooling last_token), backbone đóng băng · fusion [u; v; |u−v|; u·v], head MLP 512 · CE · lr head 0.001, 30 epoch · train thật, nhãn m1 Qwen/Qwen3-Embedding-0.6B0.227m1 93.87% ±0.5383.06% ±0.5659.33% ±2.5393.41% ±1.4276.01% ±4.0077.23% ±4.1792.38% ±0.6581.77% ±0.4958.54% ±2.5693.34% ±1.4476.18% ±4.0744.34% ±15.71
115 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-IDF0.282m1 94.91%85.71%59.89%92.54%85.37%76.92%93.74%84.50%59.48%92.46%85.71%57.14%
116 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-IDF0.190m1 95.67%86.69%66.26%93.00%85.37%69.57%94.57%85.50%65.58%92.93%85.71%36.36%
117 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-IDF0.310m1 94.94%86.20%62.01%92.54%86.75%80.00%93.74%85.05%61.54%92.46%87.12%61.54%
118 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-IDF0.356m1 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/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_bamibert_cross_ce_real_frozenemb_lr2e5_m1 ––m1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_bamibert_cross_ce_real_frozenemb_lr4e5_m1 ––m1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e3_fbverifier_devtop1_familygeomean_fbf1_m1fallback_verifier · train thật, nhãn m1 –hardm1 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_m1fallback_verifier · train thật, nhãn m1 –hardm1 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_allensemble trung bình trên 23 phương pháp · nhãn v1 –0.065v1 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_allensemble chọn tham lam Caruana trên 23 phương pháp · nhãn v1 –0.070v1 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_lr4e5XLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.023v1 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-v20.011v1 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_frozenembXLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE trọng số căn nghịch tần suất · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.021v1 94.87% ±0.1088.39% ±0.5465.20% ±1.7492.36% ±1.2785.05% ±1.4253.03% ±18.3793.42% ±0.2487.44% ±0.4264.39% ±1.1292.89% ±1.3085.44% ±1.4542.96% ±22.37
6 e3_prior_em_mmbert_bottypeEM điều chỉnh prior trên e2_mmbert_base_cross_ce_real_frozenemb_lr2e5, phạm vi theo loại câu bot (transductive) · nhãn v1 –0.082v1 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_globalEM điều chỉnh prior trên e2_mmbert_base_cross_ce_real_frozenemb_lr2e5, phạm vi toàn bộ (transductive) · nhãn v1 –0.082v1 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_lr2e5mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 jhu-clsp/mmBERT-base0.061v1 95.32% ±0.1488.07% ±1.3063.86% ±2.8491.05% ±1.6785.88% ±2.8674.24% ±8.4494.18% ±0.2287.17% ±1.5063.20% ±3.1191.57% ±1.7186.29% ±2.9268.89% ±10.18
9 e2_xlmr_base_cross_ce_real_frozenembXLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.035v1 95.14% ±0.1788.61% ±0.7667.48% ±0.9892.06% ±0.4084.15% ±1.0755.79% ±10.3193.85% ±0.1787.74% ±0.8566.95% ±1.3392.59% ±0.4084.53% ±1.0945.98% ±10.47
10 e2_xlmr_base_cross_logitadjust_real_frozenembXLM-R base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE điều chỉnh logit theo prior (τ=1) · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.035v1 95.04% ±0.4488.03% ±0.4063.79% ±3.1992.73% ±1.6986.46% ±1.9046.34% ±3.5793.65% ±0.5387.09% ±0.4662.82% ±2.5093.28% ±1.7386.89% ±1.9437.52% ±4.87
11 e2_qwen3_06b_base_cross_ce_real_loraQwen3-0.6B Base · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn v1 Qwen/Qwen3-0.6B-Base0.098v1 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_allensemble stacking LR trên 23 phương pháp · nhãn v1 –0.088v1 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_loraQwen3-1.7B Base · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0001, 3 epoch · train thật, nhãn v1 Qwen/Qwen3-1.7B-Base0.083v1 94.96% ±0.0987.79% ±0.3562.72% ±1.6892.81% ±0.3586.86% ±1.6469.46% ±12.9293.67% ±0.0886.84% ±0.4562.11% ±1.9093.37% ±0.3687.30% ±1.6763.30% ±16.33
14 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 vinai/phobert-base-v20.022v1 95.22% ±0.1688.54% ±0.6766.58% ±0.8392.71% ±1.2784.91% ±2.4658.33% ±14.4393.96% ±0.2487.59% ±0.7565.99% ±1.1193.26% ±1.3185.30% ±2.5244.44% ±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-base0.088v1 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_lr4e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 intfloat/multilingual-e5-base0.028v1 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_frozenembXLM-R base · chỉ câu KH · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 FacebookAI/xlm-roberta-base0.076v1 94.59% ±0.3487.90% ±1.4366.84% ±2.5791.68% ±0.8186.38% ±1.6255.72% ±14.9993.13% ±0.5086.90% ±1.6566.16% ±2.5992.20% ±0.8386.81% ±1.6542.86% ±20.76
18 e2_me5_base_cross_ce_real_frozenemb_lr2e5multilingual-E5 base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 intfloat/multilingual-e5-base0.056v1 95.02% ±0.2388.98% ±0.3466.67% ±1.7591.07% ±1.0285.04% ±1.7650.37% ±14.2993.65% ±0.2588.25% ±0.4165.50% ±1.8191.57% ±1.0485.45% ±1.8036.90% ±17.62
19 e2_gemma3_270m_cross_ce_real_loragemma-3-270m · câu bot + câu KH ghép cặp · LoRA r=16 · CE · lr 0.0002, 3 epoch · train thật, nhãn v1 unsloth/gemma-3-270m0.116v1 94.61% ±0.4387.68% ±0.5362.98% ±2.1090.62% ±0.9582.40% ±1.2651.52% ±7.6193.19% ±0.6086.47% ±0.5962.67% ±2.8291.34% ±0.9882.95% ±1.0142.59% ±8.49
20 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozenQwen3-Embedding-0.6B · cross-attention (Q = câu KH, K/V = câu bot), backbone đóng băng · CE trọng số căn nghịch tần suất, label smoothing 0.05 · lr 0.0003, 40 epoch · train thật, nhãn v1 Qwen/Qwen3-Embedding-0.6B0.114v1 94.29% ±0.1486.13% ±0.4961.02% ±1.1788.70% ±0.8387.42% ±2.0062.22% ±7.7092.78% ±0.3184.91% ±0.5659.98% ±0.9089.15% ±0.8687.87% ±2.0555.38% ±7.99
21 e2_mmbert_base_cross_ce_real_frozenemb_lr4e5mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 jhu-clsp/mmBERT-base0.075v1 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_lr1e5PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 vinai/phobert-base-v20.026v1 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-base0.123v1 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_lr1e5mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 jhu-clsp/mmBERT-base0.120v1 94.87% ±0.1688.28% ±1.0865.42% ±0.2493.43% ±1.4485.49% ±0.8545.00% ±7.0793.77% ±0.0687.51% ±1.2065.62% ±0.1794.00% ±1.4786.18% ±1.2635.38% ±6.53
25 e2_visobert_cross_ce_real_frozenemb_lr2e5ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 uitnlp/visobert0.093v1 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_lr4e5ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 uitnlp/visobert0.066v1 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-IDF0.123v1 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-IDF0.256v1 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-IDF0.125v1 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_lr1e5ViSoBERT · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 uitnlp/visobert0.232v1 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-IDF0.234v1 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/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_mdeberta_cross_ce_real_frozenemb_lr2e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 ––v1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_mdeberta_cross_ce_real_frozenemb_lr2e5_fp32mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 ––v1 N/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_mdeberta_cross_ce_real_frozenemb_lr4e5mDeBERTa-v3 · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 microsoft/mdeberta-v3-baseargmaxv1 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_lr1e5ViDeBERTa · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 1e-05, 4 epoch · train thật, nhãn v1 Fsoft-AIC/videberta-baseargmaxv1 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_lr2e5ViDeBERTa · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v1 Fsoft-AIC/videberta-baseargmaxv1 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/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/AN/A
– e2_videberta_cross_ce_real_frozenemb_lr4e5ViDeBERTa · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 4e-05, 4 epoch · train thật, nhãn v1 Fsoft-AIC/videberta-baseargmaxv1 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_lv21mmBERT base · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v2_1 jhu-clsp/mmBERT-base0.076v2_1 94.92% ±0.3688.37% ±0.8364.36% ±3.0592.42% ±0.7686.37% ±3.0460.85% ±14.4093.78% ±0.3787.95% ±0.7664.92% ±2.6292.98% ±0.7886.80% ±3.0950.79% ±19.83
2 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_lv21PhoBERT-v2 base (tách từ underthesea) · câu bot + câu KH ghép cặp · full fine-tune, đóng băng embedding · CE · lr 2e-05, 4 epoch · train thật, nhãn v2_1 vinai/phobert-base-v20.033v2_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.

Precision theo từng intent (dev và test_full)

Recall theo từng intent (dev và test_full)

F1 theo từng intent (dev và test_full)

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.

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.

STTModel_NameData clean seed · dòng ASR/chép tayclean Macro_F1 ASRclean Macro_F1 chép tayclean Macro_F1 chênhclean Macro_F1 −rep ASRclean Macro_F1 −rep chép tayclean Macro_F1 −rep chênhclean FB recall ASRclean FB recall chép tayclean FB recall chênhclean FB precision ASRclean FB precision chép tayclean FB precision chênhreply-unseen seed · dòng ASR/chép tayreply-unseen Macro_F1 ASRreply-unseen Macro_F1 chép tayreply-unseen Macro_F1 chênhreply-unseen Macro_F1 −rep ASRreply-unseen Macro_F1 −rep chép tayreply-unseen Macro_F1 −rep chênhreply-unseen FB recall ASRreply-unseen FB recall chép tayreply-unseen FB recall chênhreply-unseen FB precision ASRreply-unseen FB precision chép tayreply-unseen FB precision chênh
14 e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m2m2 3 · 2310.0/2310.086.69%89.52% ±1.05+2.889.06%91.35% ±1.86+2.396.00%94.40% ±0.80-1.653.41%55.17% ±7.77+1.83 · 1894.0/1894.080.95%84.64% ±2.73+3.788.81%91.22% ±1.88+2.496.58%94.87% ±0.85-1.752.82%54.63% ±7.52+1.8
15 e2_me5_base_cross_ce_real_frozenemb_lr4e5_m2m2 1 · 2310.0/2310.083.62%88.23%+4.687.49%90.88%+3.496.00%96.00%+0.049.38%54.79%+5.41 · 1894.0/1894.078.71%83.77%+5.186.95%90.53%+3.695.73%95.73%+0.048.07%53.59%+5.5
16 e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m2m2 1 · 2310.0/2310.087.41%90.23%+2.889.51%92.28%+2.888.00%87.20%-0.863.58%68.99%+5.41 · 1894.0/1894.083.83%87.00%+3.289.17%92.09%+2.988.03%87.18%-0.962.42%68.00%+5.6
17 e2_qwen3_17b_base_genlabel_ce_real_lora_m2m2 1 · 2310.0/2310.087.74%90.58%+2.889.57%92.98%+3.493.60%94.40%+0.855.45%62.11%+6.71 · 1894.0/1894.084.74%87.73%+3.089.19%92.78%+3.693.16%94.02%+0.954.23%61.11%+6.9
18 e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m2m2 3 · 2310.0/2310.089.45%91.59% ±0.25+2.189.96%92.46% ±0.44+2.595.47%94.13% ±1.22-1.354.82%59.94% ±0.81+5.13 · 1894.0/1894.087.17%89.58% ±0.29+2.489.59%92.24% ±0.45+2.695.16%93.73% ±1.31-1.453.87%59.19% ±1.13+5.3
20 e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m2m2 3 · 2310.0/2310.087.94%91.04% ±0.42+3.190.26%93.20% ±1.48+2.996.27%95.73% ±0.46-0.559.78%67.09% ±10.16+7.33 · 1894.0/1894.084.46%87.98% ±0.24+3.590.01%93.14% ±1.51+3.196.58%96.01% ±0.49-0.658.92%66.52% ±10.10+7.6
21 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m2m2 3 · 2310.0/2310.086.25%88.97% ±3.04+2.788.98%92.13% ±1.86+3.195.73%95.20% ±2.12-0.554.90%60.22% ±10.30+5.33 · 1894.0/1894.080.85%83.63% ±5.45+2.888.68%92.00% ±2.01+3.396.01%95.44% ±2.15-0.654.16%59.70% ±10.85+5.5
22 e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m2m2 3 · 2310.0/2310.087.97%91.46% ±0.09+3.590.69%93.86% ±0.42+3.296.27%96.53% ±0.46+0.357.52%64.77% ±0.90+7.23 · 1894.0/1894.084.02%88.48% ±0.41+4.590.38%93.82% ±0.47+3.496.01%96.30% ±0.49+0.356.86%64.75% ±0.50+7.9
23 e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m2m2 2 · 2310.0/2310.087.08%89.99% ±0.35+2.989.12%91.92% ±1.10+2.896.40%95.60% ±0.57-0.853.23%60.07% ±1.35+6.82 · 1894.0/1894.082.88%86.25% ±1.01+3.488.84%91.83% ±1.14+3.097.01%96.15% ±0.60-0.952.45%59.54% ±1.41+7.1
24 e2_gemma3_1b_pt_cross_ce_real_lora_r32_m2m2 3 · 2310.0/2310.086.15%89.06% ±2.87+2.988.26%90.55% ±1.93+2.395.47%94.67% ±0.46-0.851.30%55.15% ±6.06+3.83 · 1894.0/1894.082.34%86.17% ±4.69+3.887.95%90.45% ±2.07+2.596.01%95.16% ±0.49-0.950.48%54.63% ±6.32+4.1
25 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m2m2 3 · 2310.0/2310.085.65%88.14% ±0.55+2.588.22%91.02% ±0.49+2.892.80%91.47% ±1.67-1.354.21%59.05% ±3.93+4.83 · 1894.0/1894.082.70%85.22% ±0.49+2.588.10%91.05% ±0.48+2.992.88%91.45% ±1.48-1.454.50%59.57% ±2.96+5.1
26 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m2m2 1 · 2310.0/2310.079.10%81.36%+2.389.21%91.92%+2.795.20%95.20%+0.056.13%59.50%+3.41 · 1894.0/1894.079.25%81.65%+2.489.39%92.27%+2.996.58%96.58%+0.057.36%61.08%+3.7
27 e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m2m2 1 · 2310.0/2310.088.56%91.24%+2.789.73%92.94%+3.297.60%97.60%+0.054.95%61.00%+6.01 · 1894.0/1894.086.26%89.10%+2.889.39%92.80%+3.497.44%97.44%+0.054.03%60.32%+6.3
28 e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m2m2 1 · 2310.0/2310.089.89%92.84%+3.091.57%95.11%+3.594.40%95.20%+0.866.29%76.28%+10.01 · 1894.0/1894.087.25%90.32%+3.191.36%95.05%+3.794.02%94.87%+0.965.48%75.51%+10.0
29 e3_ensemble_seed6_kd_a05_t1_m2m2 1 · 2310.0/2310.089.54%91.94%+2.491.09%93.19%+2.196.80%96.80%+0.060.80%63.68%+2.91 · 1894.0/1894.085.79%88.76%+3.090.95%93.17%+2.297.44%97.44%+0.060.32%63.33%+3.0
30 e2_me5_base_cross_ce_real_frozenemb_lr2e5_m2m2 3 · 2310.0/2310.085.51%88.61% ±0.82+3.189.10%91.97% ±1.30+2.996.53%96.27% ±1.67-0.355.47%60.50% ±6.69+5.03 · 1894.0/1894.079.06%82.27% ±0.39+3.288.69%91.72% ±1.34+3.096.30%96.01% ±1.78-0.354.39%59.59% ±6.86+5.2
32 e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m2m2 3 · 2310.0/2310.089.41%91.95% ±0.53+2.591.57%94.53% ±1.39+3.095.73%97.07% ±1.22+1.363.85%70.35% ±10.85+6.53 · 1894.0/1894.086.46%89.09% ±0.99+2.691.35%94.54% ±1.52+3.295.44%96.87% ±1.31+1.463.18%70.25% ±11.51+7.1
33 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m2m2 3 · 2310.0/2310.087.11%88.84% ±1.53+1.789.20%90.96% ±0.76+1.894.13%93.87% ±3.03-0.354.24%55.56% ±2.59+1.33 · 1894.0/1894.082.44%84.73% ±2.92+2.388.93%90.81% ±0.72+1.994.02%93.73% ±2.75-0.353.56%54.94% ±2.31+1.4
34 e2_gemma3_1b_pt_cross_ce_real_lora_m2m2 3 · 2310.0/2310.086.52%89.38% ±1.61+2.989.19%91.66% ±0.91+2.595.20%93.60% ±0.80-1.655.38%60.22% ±3.06+4.83 · 1894.0/1894.082.19%85.80% ±2.56+3.689.06%91.70% ±0.84+2.696.01%94.30% ±1.31-1.755.26%60.38% ±2.59+5.1
35 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m2m2 3 · 2310.0/2310.088.44%90.37% ±0.85+1.990.29%92.30% ±0.51+2.093.60%92.80% ±1.60-0.861.33%63.83% ±2.61+2.53 · 1894.0/1894.084.82%87.13% ±2.73+2.390.15%92.32% ±0.62+2.294.02%93.16% ±1.71-0.960.99%63.64% ±3.16+2.6
36 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m2m2 3 · 2310.0/2310.085.57%88.13% ±1.47+2.688.54%91.44% ±1.59+2.993.60%93.33% ±3.33-0.356.84%60.48% ±8.68+3.63 · 1894.0/1894.081.45%84.18% ±1.95+2.788.22%91.27% ±1.65+3.193.73%93.45% ±3.24-0.356.07%59.87% ±8.80+3.8
37 e2_qwen3_17b_base_cross_ce_real_lora_m2m2 1 · 2310.0/2310.088.71%91.06%+2.390.74%92.98%+2.290.40%90.40%+0.060.43%68.07%+7.61 · 1894.0/1894.085.85%88.56%+2.790.52%92.94%+2.490.60%90.60%+0.059.89%67.95%+8.1
38 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m2m2 6 · 2310.0/2310.088.46%90.45% ±1.77+2.090.23%92.26% ±1.06+2.095.87%95.07% ±1.63-0.858.47%61.30% ±5.43+2.86 · 1894.0/1894.084.46%87.04% ±3.03+2.690.10%92.26% ±1.07+2.296.30%95.44% ±1.85-0.958.21%61.20% ±5.61+3.0
39 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m2m2 3 · 2310.0/2310.087.35%89.29% ±1.54+1.989.67%91.64% ±1.16+2.095.47%95.20% ±0.00-0.355.84%57.79% ±5.22+2.03 · 1894.0/1894.082.70%85.21% ±2.89+2.589.41%91.51% ±1.25+2.195.44%95.16% ±0.49-0.355.31%57.37% ±5.72+2.1
40 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m2m2 3 · 2310.0/2310.087.13%89.62% ±1.85+2.589.30%91.47% ±0.85+2.294.40%94.93% ±0.92+0.556.84%58.41% ±3.26+1.63 · 1894.0/1894.083.23%86.12% ±3.65+2.989.08%91.39% ±0.93+2.394.87%95.16% ±0.49+0.356.28%57.94% ±3.51+1.7
41 e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m2m2 3 · 2310.0/2310.086.31%88.84% ±0.28+2.588.63%91.29% ±1.29+2.794.40%93.60% ±1.39-0.853.03%58.70% ±4.81+5.73 · 1894.0/1894.082.49%85.48% ±0.36+3.088.41%91.26% ±1.13+2.995.16%94.30% ±2.47-0.952.42%58.33% ±4.15+5.9
42 e2_gemma3_270m_cross_ce_real_lora_m2m2 3 · 2310.0/2310.083.76%86.34% ±0.93+2.686.89%89.50% ±1.38+2.693.07%91.73% ±2.01-1.351.17%54.68% ±3.27+3.53 · 1894.0/1894.079.75%82.65% ±1.81+2.986.56%89.33% ±1.54+2.893.73%92.31% ±1.71-1.450.72%54.39% ±3.73+3.7
43 e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m2m2 3 · 2310.0/2310.087.09%89.62% ±1.08+2.589.85%92.45% ±0.41+2.694.67%93.33% ±0.92-1.357.94%63.92% ±1.86+6.03 · 1894.0/1894.082.40%85.21% ±1.77+2.889.69%92.44% ±0.40+2.795.44%94.02% ±0.85-1.457.30%63.53% ±2.30+6.2
44 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m2m2 3 · 2310.0/2310.084.39%87.28% ±2.04+2.987.69%90.48% ±2.45+2.895.47%95.47% ±1.67+0.050.75%54.79% ±5.99+4.03 · 1894.0/1894.079.33%82.40% ±1.48+3.187.35%90.28% ±2.37+2.995.44%95.44% ±1.31+0.049.87%54.04% ±5.65+4.2
45 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m2m2 3 · 2310.0/2310.086.11%88.08% ±1.45+2.089.10%91.78% ±0.57+2.795.73%94.93% ±2.01-0.854.48%59.08% ±4.02+4.63 · 1894.0/1894.082.67%84.73% ±1.91+2.188.92%91.77% ±0.47+2.996.01%95.16% ±1.78-0.954.35%59.20% ±2.37+4.9
46 e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m2m2 1 · 2310.0/2310.088.31%90.85%+2.589.68%92.73%+3.094.40%92.80%-1.659.90%66.29%+6.41 · 1894.0/1894.085.75%88.40%+2.689.57%92.74%+3.294.87%93.16%-1.760.00%66.46%+6.5
47 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m2m2 3 · 2310.0/2310.085.71%88.54% ±1.26+2.888.56%91.39% ±0.66+2.893.07%93.60% ±1.39+0.554.56%59.02% ±2.96+4.53 · 1894.0/1894.081.28%84.31% ±2.55+3.088.31%91.32% ±0.71+3.093.45%94.02% ±1.48+0.654.08%58.78% ±3.48+4.7
48 e2_xlmr_base_khonly_ce_real_frozenemb_m2m2 1 · 2310.0/2310.084.89%87.80%+2.987.96%90.81%+2.996.00%96.00%+0.052.40%56.87%+4.51 · 1894.0/1894.079.08%82.21%+3.187.62%90.65%+3.096.58%96.58%+0.051.60%56.22%+4.6
49 e3_ensemble_seed_kd_a05_t1_m2m2 1 · 2310.0/2310.089.62%92.08%+2.591.18%93.10%+1.996.80%96.00%-0.860.50%63.83%+3.31 · 1894.0/1894.085.93%89.76%+3.891.11%93.17%+2.197.44%96.58%-0.960.32%63.84%+3.5
50 e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m2m2 3 · 2310.0/2310.086.92%89.39% ±0.84+2.587.99%90.72% ±0.09+2.794.93%94.40% ±2.12-0.551.46%55.46% ±2.95+4.03 · 1894.0/1894.083.51%86.23% ±1.89+2.787.81%90.71% ±0.07+2.995.44%94.87% ±2.26-0.651.37%55.63% ±2.67+4.3
51 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m2m2 3 · 2310.0/2310.087.11%89.33% ±1.41+2.288.80%91.10% ±0.46+2.396.00%96.27% ±0.46+0.351.89%54.71% ±1.05+2.83 · 1894.0/1894.082.52%85.38% ±2.88+2.988.54%90.94% ±0.46+2.496.01%96.30% ±0.49+0.351.31%54.09% ±0.80+2.8
52 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m2m2 3 · 2310.0/2310.086.63%88.63% ±1.82+2.088.83%90.57% ±1.46+1.793.60%93.07% ±2.81-0.554.30%55.11% ±5.31+0.83 · 1894.0/1894.081.94%84.33% ±3.64+2.488.55%90.43% ±1.43+1.994.02%93.45% ±3.00-0.653.41%54.31% ±5.02+0.9
53 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t2_tsa_m2m2 3 · 2310.0/2310.089.02%90.32% ±0.46+1.389.73%91.98% ±0.49+2.293.60%93.07% ±1.67-0.557.43%59.90% ±4.23+2.53 · 1894.0/1894.086.62%87.63% ±0.89+1.089.52%91.85% ±0.63+2.394.59%93.45% ±1.78-1.157.06%59.49% ±4.64+2.4
54 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m2m2 3 · 2310.0/2310.087.64%89.64% ±1.53+2.089.31%91.48% ±0.89+2.296.00%96.27% ±0.92+0.356.35%58.24% ±4.89+1.93 · 1894.0/1894.083.49%85.67% ±3.04+2.289.08%91.36% ±0.98+2.396.30%96.58% ±0.85+0.355.69%57.67% ±5.07+2.0
55 e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m2m2 3 · 2310.0/2310.082.91%85.66% ±4.78+2.889.02%91.51% ±1.34+2.592.00%90.67% ±0.92-1.355.37%60.69% ±6.97+5.33 · 1894.0/1894.079.55%83.12% ±3.10+3.688.93%91.61% ±1.46+2.792.31%90.88% ±0.99-1.455.75%61.52% ±7.73+5.8
56 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m2m2 3 · 2310.0/2310.086.68%89.31% ±0.59+2.688.76%91.51% ±0.53+2.796.00%94.67% ±1.22-1.352.92%57.88% ±4.57+5.03 · 1894.0/1894.082.90%85.82% ±1.20+2.988.53%91.44% ±0.50+2.997.15%95.73% ±0.85-1.452.25%57.37% ±4.28+5.1
57 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m2m2 3 · 2310.0/2310.086.45%89.17% ±1.05+2.789.71%92.73% ±0.49+3.093.87%93.87% ±1.85+0.056.83%63.00% ±1.56+6.23 · 1894.0/1894.078.85%81.74% ±1.05+2.989.57%92.80% ±0.55+3.294.87%94.87% ±2.26+0.056.50%62.99% ±1.82+6.5
58 e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m2m2 3 · 2310.0/2310.087.28%90.31% ±2.11+3.090.21%93.07% ±1.37+2.992.00%92.27% ±0.92+0.361.41%66.47% ±7.43+5.13 · 1894.0/1894.081.58%85.21% ±4.37+3.690.11%93.17% ±1.31+3.193.16%93.45% ±0.99+0.360.97%66.38% ±7.04+5.4
59 e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m2m2 3 · 2310.0/2310.086.90%89.00% ±1.04+2.189.13%91.15% ±0.87+2.094.40%94.40% ±1.39+0.055.44%56.76% ±2.80+1.33 · 1894.0/1894.082.87%85.43% ±1.98+2.688.92%91.11% ±0.97+2.294.59%94.59% ±0.99+0.055.05%56.62% ±3.30+1.6
60 e2_xlmr_base_cross_logitadjust_real_frozenemb_m2m2 3 · 2310.0/2310.084.70%88.11% ±0.98+3.487.42%90.72% ±0.98+3.395.20%95.47% ±0.92+0.349.47%53.71% ±3.12+4.23 · 1894.0/1894.080.07%83.85% ±1.18+3.887.08%90.58% ±1.15+3.595.73%96.01% ±1.78+0.348.88%53.30% ±3.62+4.4
61 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m2m2 3 · 2310.0/2310.088.12%91.06% ±0.74+2.990.93%93.47% ±0.79+2.594.13%93.07% ±1.85-1.162.23%67.66% ±6.03+5.43 · 1894.0/1894.084.44%87.81% ±0.77+3.490.66%93.35% ±0.76+2.794.02%92.88% ±2.47-1.161.31%66.96% ±5.93+5.7
62 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m2m2 3 · 2310.0/2310.085.55%87.90% ±1.17+2.388.79%91.47% ±0.40+2.790.93%90.13% ±4.89-0.858.33%63.07% ±4.43+4.73 · 1894.0/1894.080.96%83.36% ±2.30+2.488.48%91.31% ±0.46+2.891.17%90.31% ±5.22-0.957.47%62.39% ±4.26+4.9
63 e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m2m2 3 · 2310.0/2310.085.38%87.79% ±0.61+2.488.80%91.54% ±0.34+2.794.67%92.80% ±1.60-1.956.47%62.18% ±1.81+5.73 · 1894.0/1894.080.99%83.49% ±1.62+2.588.52%91.44% ±0.47+2.995.16%93.16% ±1.71-2.055.89%61.87% ±2.11+6.0
64 e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m2m2 1 · 2310.0/2310.086.20%88.35%+2.187.44%90.02%+2.693.60%94.40%+0.847.18%51.98%+4.81 · 1894.0/1894.082.86%85.15%+2.387.12%89.87%+2.794.02%94.87%+0.946.61%51.63%+5.0
65 e2_visobert_cross_ce_real_frozenemb_lr4e5_m2m2 1 · 2310.0/2310.082.82%84.62%+1.884.98%87.14%+2.294.40%92.80%-1.640.83%42.65%+1.81 · 1894.0/1894.078.96%80.89%+1.984.76%87.07%+2.394.87%93.16%-1.740.96%42.91%+2.0
66 e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m2m2 1 · 2310.0/2310.085.11%88.95%+3.887.73%91.09%+3.490.40%91.20%+0.849.56%54.55%+5.01 · 1894.0/1894.080.64%85.03%+4.487.54%91.12%+3.691.45%92.31%+0.949.54%54.82%+5.3
67 e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m2m2 3 · 2310.0/2310.086.97%88.88% ±1.27+1.989.15%91.45% ±1.44+2.396.00%94.93% ±1.22-1.153.35%57.74% ±6.67+4.43 · 1894.0/1894.081.27%83.31% ±2.36+2.089.02%91.47% ±1.26+2.496.58%95.44% ±1.31-1.153.24%57.88% ±5.56+4.6
68 e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m2m2 3 · 2310.0/2310.087.96%90.06% ±0.66+2.190.15%92.46% ±0.39+2.392.53%92.53% ±1.85+0.060.55%62.61% ±3.92+2.13 · 1894.0/1894.083.35%85.61% ±2.17+2.389.82%92.29% ±0.55+2.592.88%92.88% ±1.97+0.059.96%62.09% ±4.19+2.1
69 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m2m2 3 · 2310.0/2310.085.26%88.21% ±1.64+2.988.49%91.14% ±1.10+2.693.60%93.33% ±1.22-0.354.35%58.34% ±4.05+4.03 · 1894.0/1894.080.45%83.51% ±3.25+3.188.13%90.92% ±1.17+2.893.73%93.45% ±1.78-0.353.16%57.25% ±4.00+4.1
70 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m2m2 3 · 2310.0/2310.087.03%89.98% ±1.68+2.989.72%92.68% ±0.82+3.093.60%93.07% ±2.01-0.561.36%66.46% ±6.42+5.13 · 1894.0/1894.083.14%86.39% ±2.40+3.389.47%92.59% ±0.81+3.194.02%93.45% ±1.78-0.660.52%65.81% ±6.36+5.3
71 e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m2m2 3 · 2310.0/2310.086.68%89.57% ±1.82+2.989.26%92.06% ±2.15+2.896.00%96.80% ±0.80+0.853.36%59.46% ±8.39+6.13 · 1894.0/1894.083.32%86.77% ±2.52+3.488.96%91.95% ±2.31+3.096.01%96.87% ±0.99+0.952.78%59.19% ±8.74+6.4
72 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t1_tsa_m2m2 3 · 2310.0/2310.088.47%89.80% ±1.08+1.389.40%91.58% ±0.63+2.295.47%94.67% ±2.44-0.856.35%58.68% ±3.61+2.33 · 1894.0/1894.085.64%86.42% ±2.92+0.889.14%91.45% ±0.73+2.395.73%94.87% ±2.26-0.955.81%58.24% ±3.95+2.4
73 e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m2m2 3 · 2310.0/2310.086.20%87.77% ±2.25+1.687.68%89.56% ±0.96+1.992.27%91.73% ±4.03-0.547.85%50.40% ±4.51+2.63 · 1894.0/1894.082.69%84.43% ±4.38+1.787.85%89.94% ±1.53+2.192.59%92.02% ±4.30-0.650.09%53.25% ±7.62+3.2
74 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m2m2 3 · 2310.0/2310.088.10%89.60% ±1.18+1.589.93%92.00% ±2.10+2.193.33%92.80% ±2.88-0.561.64%63.02% ±8.88+1.43 · 1894.0/1894.084.08%86.03% ±1.38+1.989.65%91.87% ±2.12+2.292.88%92.31% ±3.08-0.661.18%62.63% ±9.01+1.4
75 e2_xlmr_base_cross_ce_real_frozenemb_m2m2 3 · 2310.0/2310.084.79%88.15% ±1.89+3.487.92%90.93% ±1.43+3.091.73%93.07% ±3.70+1.353.13%57.86% ±8.71+4.73 · 1894.0/1894.079.16%82.75% ±3.99+3.687.68%90.86% ±1.61+3.292.59%94.02% ±3.73+1.452.71%57.65% ±9.32+4.9
76 e2_xlmr_base_cross_wcesqrt_real_frozenemb_m2m2 3 · 2310.0/2310.086.07%89.23% ±0.93+3.288.29%91.51% ±1.04+3.294.40%94.13% ±2.31-0.354.14%59.16% ±6.71+5.03 · 1894.0/1894.082.31%85.77% ±1.47+3.587.92%91.34% ±1.18+3.495.16%94.87% ±1.48-0.353.26%58.43% ±6.85+5.2
77 e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m2m2 3 · 2310.0/2310.086.10%87.24% ±1.72+1.188.98%90.64% ±0.49+1.795.20%94.67% ±0.46-0.554.85%55.30% ±0.56+0.43 · 1894.0/1894.081.79%83.04% ±2.70+1.388.78%90.53% ±0.49+1.895.44%94.59% ±0.99-0.954.50%54.98% ±0.87+0.5
78 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m2m2 3 · 2310.0/2310.085.07%86.26% ±4.57+1.289.40%91.24% ±0.98+1.892.80%92.80% ±2.12+0.059.12%59.47% ±7.73+0.43 · 1894.0/1894.082.81%83.96% ±2.01+1.189.27%91.20% ±0.91+1.993.45%92.59% ±2.61-0.958.97%59.38% ±7.20+0.4
79 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsa_m2m2 3 · 2310.0/2310.087.73%89.30% ±1.17+1.690.07%92.14% ±1.23+2.194.13%93.60% ±2.12-0.559.44%61.90% ±5.54+2.53 · 1894.0/1894.083.78%85.75% ±1.70+2.089.85%92.05% ±1.30+2.294.30%93.73% ±1.78-0.658.99%61.58% ±5.59+2.6
80 e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m2m2 2 · 2310.0/2310.083.16%85.04% ±0.70+1.985.49%87.58% ±1.84+2.192.80%92.40% ±2.83-0.450.76%53.98% ±0.49+3.22 · 1894.0/1894.079.14%81.13% ±0.32+2.084.81%87.01% ±1.96+2.292.31%91.88% ±3.02-0.449.53%52.70% ±0.09+3.2
81 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m2m2 1 · 2310.0/2310.078.16%78.95%+0.888.79%91.88%+3.190.40%91.20%+0.858.85%62.30%+3.41 · 1894.0/1894.077.79%76.82%-1.088.90%92.18%+3.391.45%92.31%+0.960.11%63.91%+3.8
82 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m2m2 3 · 2310.0/2310.083.46%86.17% ±0.51+2.786.20%88.91% ±1.16+2.786.67%86.13% ±0.46-0.555.34%57.82% ±5.50+2.53 · 1894.0/1894.079.54%82.51% ±0.56+3.085.61%88.47% ±1.30+2.986.04%85.47% ±0.00-0.654.02%56.57% ±5.88+2.5
83 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m2m2 3 · 2310.0/2310.087.43%89.08% ±0.32+1.789.57%91.31% ±1.08+1.795.47%94.93% ±2.01-0.556.20%57.73% ±6.09+1.53 · 1894.0/1894.083.36%85.22% ±0.57+1.989.26%91.12% ±1.08+1.995.73%95.16% ±2.15-0.655.26%56.83% ±5.93+1.6
84 e2_me5_base_cross_ce_real_frozenemb_lr1e5_m2m2 1 · 2310.0/2310.078.62%79.03%+0.487.68%90.12%+2.492.00%92.00%+0.051.11%55.29%+4.21 · 1894.0/1894.079.02%78.48%-0.587.55%90.18%+2.692.31%92.31%+0.051.67%56.25%+4.6
85 e2_visobert_cross_ce_real_frozenemb_lr2e5_m2m2 1 · 2310.0/2310.080.98%84.49%+3.585.52%87.47%+2.091.20%90.40%-0.843.51%45.75%+2.21 · 1894.0/1894.073.81%78.86%+5.085.24%87.36%+2.192.31%91.45%-0.943.37%45.73%+2.4
86 e2_visobert_cross_ce_real_frozenemb_lr1e5_m2m2 1 · 2310.0/2310.080.42%83.70%+3.384.20%85.89%+1.787.20%88.80%+1.645.61%47.84%+2.21 · 1894.0/1894.074.59%78.02%+3.483.79%85.62%+1.887.18%88.89%+1.745.33%47.71%+2.4
87 e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m2m2 1 · 2310.0/2310.081.36%86.07%+4.787.26%89.96%+2.791.20%92.80%+1.654.55%58.00%+3.51 · 1894.0/1894.074.69%81.49%+6.886.97%89.79%+2.892.31%94.02%+1.754.00%57.59%+3.6
88 e1_tfidf_hybrid_qwen3emb_lr_m2m2 1 · 2310.0/2310.081.85%83.54%+1.784.30%86.92%+2.685.60%86.40%+0.849.31%52.68%+3.41 · 1894.0/1894.075.98%75.94%-0.083.91%86.69%+2.885.47%86.32%+0.949.26%52.88%+3.6
90 e2_dual_qwen3emb06b_concat_ce_real_frozen_m2m2 3 · 2310.0/2310.081.27%83.41% ±1.26+2.182.08%85.16% ±1.17+3.187.20%88.80% ±0.00+1.647.17%51.32% ±5.03+4.13 · 1894.0/1894.075.26%78.03% ±1.57+2.881.44%84.67% ±1.16+3.286.89%88.03% ±0.00+1.146.39%50.55% ±4.87+4.2
14 e2_dual_bot_qwen3_06b_base_reply_gemma3_1b_pt_concat_ce_real_lora_m1m1 3 · 2358.0/2358.084.53%86.70% ±1.32+2.287.32%89.13% ±1.64+1.894.62%92.56% ±0.89-2.151.43%52.56% ±6.96+1.13 · 1942.0/1942.078.50%81.22% ±1.89+2.786.97%88.89% ±1.67+1.995.08%92.90% ±0.95-2.250.78%52.02% ±6.76+1.2
15 e2_me5_base_cross_ce_real_frozenemb_lr4e5_m1m1 1 · 2358.0/2358.081.93%85.99%+4.185.46%88.19%+2.793.85%93.08%-0.846.92%51.27%+4.31 · 1942.0/1942.076.93%81.42%+4.584.82%87.70%+2.993.44%92.62%-0.845.60%50.00%+4.4
16 e2_mmbert_base_cross_ce_real_frozenemb_lr1e5_m1m1 1 · 2358.0/2358.085.29%87.51%+2.287.53%89.63%+2.186.15%84.62%-1.560.22%64.71%+4.51 · 1942.0/1942.081.46%83.94%+2.587.09%89.30%+2.286.07%84.43%-1.658.99%63.58%+4.6
17 e2_qwen3_17b_base_genlabel_ce_real_lora_m1m1 1 · 2358.0/2358.086.04%88.36%+2.387.53%90.31%+2.891.54%91.54%+0.051.97%57.49%+5.51 · 1942.0/1942.082.94%85.36%+2.487.03%89.94%+2.990.98%90.98%+0.050.45%56.06%+5.6
18 e2_qwen3_17b_base_genlabel_ce_real_lora_noeot_m1m1 3 · 2358.0/2358.087.81%89.47% ±0.31+1.787.99%89.91% ±0.46+1.993.33%91.28% ±1.18-2.151.49%55.46% ±0.84+4.03 · 1942.0/1942.085.44%87.32% ±0.37+1.987.51%89.53% ±0.50+2.092.90%90.71% ±1.25-2.250.68%54.54% ±1.29+3.9
20 e2_me5_large_instruct_cross_ce_real_frozenemb_lr3e5_m1m1 3 · 2358.0/2358.086.18%88.64% ±0.35+2.588.36%90.54% ±1.35+2.294.36%92.82% ±0.44-1.556.58%62.15% ±8.85+5.63 · 1942.0/1942.082.53%85.35% ±0.13+2.887.99%90.31% ±1.37+2.394.54%92.90% ±0.47-1.655.64%61.38% ±8.74+5.7
21 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m1m1 3 · 2358.0/2358.084.81%86.92% ±2.94+2.187.26%89.66% ±1.76+2.494.10%92.31% ±2.04-1.851.61%55.57% ±9.47+4.03 · 1942.0/1942.079.33%81.44% ±5.35+2.186.85%89.38% ±1.88+2.594.26%92.35% ±2.06-1.950.79%54.86% ±9.90+4.1
22 e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin0_m1m1 3 · 2358.0/2358.083.77%86.26% ±2.47+2.585.01%87.58% ±1.87+2.692.56%91.54% ±5.55-1.046.27%51.11% ±8.58+4.83 · 1942.0/1942.080.16%82.98% ±3.75+2.884.82%87.61% ±1.84+2.892.08%90.98% ±5.91-1.146.89%52.22% ±8.31+5.3
23 e2_qwen3_4b_base_genlabel_ce_real_qlora_noeot_m1m1 3 · 2358.0/2358.086.01%88.81% ±0.04+2.888.90%91.27% ±0.41+2.495.38%94.36% ±0.44-1.054.20%60.15% ±1.49+6.03 · 1942.0/1942.081.84%85.44% ±0.37+3.688.46%91.00% ±0.51+2.595.08%93.99% ±0.47-1.153.35%59.44% ±1.66+6.1
24 e2_gemma3_1b_pt_cross_ce_real_lora_ep5_m1m1 2 · 2358.0/2358.085.57%87.89% ±0.40+2.387.30%89.39% ±1.03+2.195.00%93.08% ±0.00-1.950.21%55.53% ±1.80+5.32 · 1942.0/1942.081.29%84.00% ±1.05+2.786.93%89.14% ±1.09+2.295.49%93.44% ±0.00-2.049.49%54.84% ±1.86+5.3
25 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepeval_m1m1 3 · 2358.0/2358.083.55%85.75% ±2.00+2.286.35%88.10% ±0.18+1.791.79%91.54% ±0.77-0.348.47%49.64% ±1.93+1.23 · 1942.0/1942.077.76%80.33% ±4.15+2.686.10%87.95% ±0.33+1.991.80%91.53% ±0.47-0.348.25%49.47% ±2.08+1.2
26 e2_gemma3_1b_pt_cross_ce_real_lora_r32_m1m1 3 · 2358.0/2358.084.63%86.85% ±2.90+2.286.62%88.11% ±1.66+1.594.10%91.79% ±0.44-2.348.73%51.24% ±5.42+2.53 · 1942.0/1942.080.72%84.06% ±4.81+3.386.23%87.92% ±1.88+1.794.54%92.35% ±0.47-2.247.68%50.67% ±5.94+3.0
27 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_ep6_m1m1 3 · 2358.0/2358.083.89%85.91% ±0.61+2.086.38%88.49% ±0.54+2.191.03%88.72% ±1.60-2.351.25%54.75% ±3.40+3.53 · 1942.0/1942.080.76%82.80% ±0.47+2.086.13%88.34% ±0.52+2.290.98%88.52% ±1.42-2.551.33%54.93% ±2.61+3.6
28 e2_mmbert_base_cross_ce_synth1x_frozenemb_lr2e5_fbmatched_m1m1 3 · 2358.0/2358.085.91%88.31% ±1.50+2.487.60%89.84% ±0.70+2.291.79%90.51% ±3.11-1.353.27%57.25% ±5.13+4.03 · 1942.0/1942.082.53%85.29% ±2.09+2.887.23%89.59% ±0.66+2.491.80%90.44% ±3.10-1.452.44%56.54% ±4.71+4.1
29 e2_mmbert_base_cross_ce_synthm1blind_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.085.21%87.25% ±1.35+2.087.05%89.32% ±0.83+2.392.82%90.00% ±4.07-2.852.13%55.90% ±6.84+3.83 · 1942.0/1942.081.54%83.71% ±2.31+2.286.64%89.03% ±0.93+2.492.90%89.89% ±4.21-3.051.28%55.15% ±7.16+3.9
30 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr4e5_m1m1 1 · 2358.0/2358.077.27%79.25%+2.087.00%89.39%+2.493.08%92.31%-0.853.07%55.56%+2.51 · 1942.0/1942.077.30%79.41%+2.187.05%89.58%+2.594.26%93.44%-0.854.25%57.00%+2.8
31 e2_qwen3_4b_base_genlabel_ce_real_lora16_entities_noeot_m1m1 1 · 2358.0/2358.086.48%88.66%+2.287.77%90.39%+2.696.15%95.38%-0.851.44%56.88%+5.41 · 1942.0/1942.083.88%86.13%+2.387.32%90.03%+2.795.90%95.08%-0.850.43%55.77%+5.3
32 e2_qwen3_8b_base_genlabel_ce_real_qlora_ep2_noeot_m1m1 1 · 2358.0/2358.087.97%90.16%+2.289.85%92.48%+2.692.31%92.31%+0.061.86%69.36%+7.51 · 1942.0/1942.085.02%87.30%+2.389.52%92.25%+2.791.80%91.80%+0.060.87%68.29%+7.4
33 e3_ensemble_seed6_kd_a05_t1_m1m1 1 · 2358.0/2358.087.74%89.77%+2.088.92%90.58%+1.794.62%93.85%-0.857.48%59.22%+1.71 · 1942.0/1942.083.88%86.45%+2.688.66%90.41%+1.895.08%94.26%-0.856.86%58.67%+1.8
34 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrepshort_m1m1 3 · 2358.0/2358.085.28%86.56% ±0.31+1.387.49%88.71% ±0.73+1.294.87%92.56% ±1.18-2.353.03%54.47% ±1.03+1.43 · 1942.0/1942.081.12%83.16% ±0.83+2.087.20%88.53% ±0.61+1.395.08%92.62% ±0.82-2.552.74%54.25% ±1.16+1.5
35 e2_gemma3_1b_pt_cross_ce_synthm1blind_lora_m1m1 3 · 2358.0/2358.085.09%87.58% ±2.10+2.587.87%89.56% ±0.84+1.793.33%91.03% ±2.91-2.354.46%58.93% ±1.87+4.53 · 1942.0/1942.080.48%84.04% ±3.12+3.687.52%89.33% ±0.88+1.893.44%90.98% ±2.84-2.553.74%58.36% ±1.69+4.6
36 e2_me5_base_cross_ce_real_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.083.55%86.21% ±0.36+2.786.93%89.31% ±1.13+2.495.13%93.85% ±2.31-1.352.23%56.28% ±6.08+4.13 · 1942.0/1942.077.04%79.76% ±0.74+2.786.38%88.89% ±1.18+2.594.81%93.44% ±2.46-1.451.03%55.16% ±6.23+4.1
38 e2_qwen3_4b_base_genlabel_ce_real_lora16_noeot_m1m1 3 · 2358.0/2358.087.26%89.27% ±0.36+2.089.57%91.90% ±1.19+2.394.36%94.62% ±1.54+0.360.20%65.69% ±9.95+5.53 · 1942.0/1942.083.97%86.05% ±0.74+2.189.25%91.75% ±1.29+2.593.99%94.26% ±1.64+0.359.47%65.30% ±10.46+5.8
39 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.084.91%86.31% ±0.94+1.487.18%88.55% ±0.76+1.492.05%90.77% ±2.66-1.351.47%52.04% ±2.64+0.63 · 1942.0/1942.079.93%81.87% ±2.11+1.986.81%88.26% ±0.70+1.591.80%90.44% ±2.37-1.450.77%51.35% ±2.42+0.6
40 e2_gemma3_1b_pt_cross_ce_real_lora_m1m1 3 · 2358.0/2358.084.98%87.29% ±1.71+2.387.35%89.15% ±0.91+1.893.85%91.28% ±1.60-2.652.69%56.09% ±2.91+3.43 · 1942.0/1942.080.59%83.58% ±2.71+3.087.14%89.04% ±0.91+1.994.54%91.80% ±2.17-2.752.60%56.19% ±2.57+3.6
41 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_kd_a05_t1_tsa_m1m1 3 · 2358.0/2358.085.65%87.28% ±0.90+1.687.26%89.01% ±0.64+1.894.62%92.56% ±1.18-2.151.62%52.62% ±0.21+1.03 · 1942.0/1942.082.28%84.06% ±0.77+1.886.90%88.74% ±0.65+1.894.81%92.62% ±0.82-2.250.96%51.99% ±0.36+1.0
42 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthr3_m1m1 3 · 2358.0/2358.086.57%88.56% ±0.49+2.087.40%89.12% ±0.38+1.792.82%92.31% ±0.77-0.553.11%54.34% ±1.97+1.23 · 1942.0/1942.083.38%85.79% ±0.89+2.487.07%88.89% ±0.50+1.892.90%92.35% ±0.95-0.552.59%53.79% ±2.27+1.2
43 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_m1m1 3 · 2358.0/2358.086.33%87.94% ±0.99+1.688.16%89.76% ±0.45+1.691.54%90.00% ±1.54-1.558.20%59.56% ±2.04+1.43 · 1942.0/1942.082.44%84.44% ±2.63+2.087.90%89.62% ±0.55+1.791.80%90.16% ±1.64-1.657.72%59.13% ±2.47+1.4
44 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_uprep2x5_m1m1 3 · 2358.0/2358.083.85%85.46% ±1.88+1.686.93%88.57% ±0.48+1.693.08%91.79% ±1.18-1.350.22%52.69% ±1.97+2.53 · 1942.0/1942.079.38%80.93% ±4.37+1.586.56%88.32% ±0.40+1.893.17%91.80% ±0.82-1.449.43%51.97% ±1.74+2.5
45 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_rdrop1_m1m1 3 · 2358.0/2358.083.60%85.60% ±1.27+2.086.67%88.90% ±1.54+2.291.79%90.51% ±3.20-1.353.54%56.15% ±7.70+2.63 · 1942.0/1942.079.32%81.43% ±1.66+2.186.22%88.58% ±1.60+2.491.80%90.44% ±3.10-1.452.68%55.36% ±7.74+2.7
46 e2_mmbert_base_cross_ce_synthv2_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.084.00%86.30% ±0.90+2.386.21%88.30% ±0.27+2.192.56%91.28% ±2.70-1.348.07%52.66% ±1.35+4.63 · 1942.0/1942.080.58%83.17% ±1.42+2.685.87%88.10% ±0.36+2.292.62%91.26% ±2.50-1.447.62%52.44% ±1.67+4.8
47 e2_qwen3_17b_base_cross_ce_real_lora_m1m1 1 · 2358.0/2358.087.19%88.76%+1.688.91%90.21%+1.388.46%88.46%+0.056.65%62.84%+6.21 · 1942.0/1942.084.21%86.12%+1.988.56%90.01%+1.488.52%88.52%+0.055.96%62.43%+6.5
48 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t1_tsa_m1m1 6 · 2358.0/2358.086.79%88.36% ±1.76+1.688.23%89.75% ±1.05+1.593.85%92.31% ±1.69-1.555.47%57.20% ±4.97+1.76 · 1942.0/1942.082.69%84.82% ±3.00+2.187.98%89.60% ±1.05+1.694.13%92.49% ±1.90-1.655.07%56.91% ±5.09+1.8
49 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr3e5_m1m1 3 · 2358.0/2358.085.40%87.07% ±1.16+1.787.52%89.16% ±1.03+1.693.33%92.31% ±0.00-1.052.32%53.78% ±4.96+1.53 · 1942.0/1942.080.57%82.77% ±2.39+2.287.13%88.87% ±1.11+1.793.17%92.08% ±0.47-1.151.65%53.18% ±5.36+1.5
50 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_lr2e5_m1m1 3 · 2358.0/2358.084.45%86.51% ±1.76+2.187.36%89.14% ±0.89+1.892.31%92.05% ±0.89-0.353.92%54.95% ±3.19+1.03 · 1942.0/1942.080.17%82.44% ±3.24+2.387.01%88.89% ±0.95+1.992.62%92.08% ±0.47-0.553.26%54.34% ±3.42+1.1
51 e2_gemma3_1b_pt_cross_ce_real_lora_lr2e4_m1m1 3 · 2358.0/2358.084.71%86.65% ±0.34+1.986.71%88.67% ±1.39+2.093.08%90.77% ±1.33-2.350.62%54.77% ±4.72+4.23 · 1942.0/1942.080.82%83.16% ±0.45+2.386.40%88.48% ±1.24+2.193.72%91.26% ±2.37-2.550.05%54.22% ±4.09+4.2
52 e2_gemma3_270m_cross_ce_real_lora_m1m1 3 · 2358.0/2358.082.03%84.02% ±0.63+2.085.10%87.01% ±1.28+1.990.77%88.72% ±1.94-2.148.54%51.39% ±2.75+2.83 · 1942.0/1942.078.01%80.35% ±1.65+2.384.77%86.80% ±1.42+2.091.26%89.07% ±1.71-2.248.16%51.13% ±3.36+3.0
53 e2_me5_large_instruct_cross_ce_real_frozenemb_lr1e5_m1m1 3 · 2358.0/2358.085.28%87.12% ±0.75+1.888.08%89.89% ±0.51+1.893.08%90.77% ±1.33-2.354.45%58.77% ±1.88+4.33 · 1942.0/1942.080.40%82.43% ±1.51+2.087.82%89.73% ±0.45+1.993.72%91.26% ±1.25-2.553.81%58.27% ±2.38+4.5
54 e2_mmbert_base_cross_ce_augboth_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.085.14%87.69% ±1.32+2.686.29%88.64% ±0.82+2.394.10%91.79% ±1.18-2.347.86%51.50% ±2.35+3.63 · 1942.0/1942.082.83%85.76% ±1.32+2.986.44%88.99% ±0.31+2.593.99%91.80% ±1.64-2.249.72%54.13% ±0.86+4.4
55 e2_mmbert_base_cross_ce_augfiller_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.083.95%86.24% ±2.47+2.386.35%88.39% ±0.56+2.093.08%91.03% ±1.94-2.149.95%53.98% ±1.11+4.03 · 1942.0/1942.081.02%83.55% ±3.43+2.586.24%88.43% ±0.70+2.293.72%91.53% ±2.06-2.250.45%54.84% ±0.71+4.4
56 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_msd5_m1m1 3 · 2358.0/2358.082.79%85.12% ±1.71+2.385.97%88.09% ±2.33+2.194.10%92.82% ±1.18-1.347.82%51.02% ±5.20+3.23 · 1942.0/1942.077.64%80.09% ±1.19+2.585.51%87.75% ±2.24+2.293.99%92.62% ±0.82-1.446.88%50.15% ±4.86+3.3
57 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_m1m1 3 · 2358.0/2358.084.47%85.95% ±1.42+1.587.13%89.23% ±0.41+2.193.59%92.05% ±1.94-1.551.27%55.09% ±3.53+3.83 · 1942.0/1942.080.92%82.46% ±1.90+1.586.82%89.04% ±0.36+2.293.72%92.08% ±1.71-1.650.96%54.96% ±2.11+4.0
58 e2_mmbert_base_cross_ce_real_frozenemb_lr4e5_m1m1 1 · 2358.0/2358.085.72%88.21%+2.587.69%90.13%+2.492.31%90.00%-2.356.60%61.58%+5.01 · 1942.0/1942.082.69%85.39%+2.787.46%89.97%+2.592.62%90.16%-2.556.50%61.45%+5.0
59 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.084.20%86.43% ±1.26+2.286.75%88.87% ±0.60+2.191.28%90.77% ±1.33-0.550.94%54.69% ±2.02+3.83 · 1942.0/1942.079.67%82.07% ±2.53+2.486.38%88.63% ±0.64+2.291.53%90.98% ±1.42-0.550.31%54.23% ±2.32+3.9
60 e2_qwen3_17b_base_genlabel_ce_lora_noeot_synthr3_m1m1 3 · 2358.0/2358.087.34%89.48% ±1.48+2.188.53%90.43% ±1.05+1.991.28%91.03% ±1.60-0.354.11%58.80% ±3.98+4.73 · 1942.0/1942.084.40%85.68% ±3.76+1.388.07%90.07% ±1.17+2.090.98%90.71% ±1.25-0.352.95%57.59% ±4.38+4.6
61 e2_xlmr_base_khonly_ce_real_frozenemb_m1m1 1 · 2358.0/2358.083.09%85.23%+2.186.38%88.36%+2.094.62%93.08%-1.550.62%53.54%+2.91 · 1942.0/1942.076.70%78.93%+2.285.89%88.05%+2.295.08%93.44%-1.649.57%52.78%+3.2
62 e3_ensemble_seed_kd_a05_t1_m1m1 1 · 2358.0/2358.087.51%89.39%+1.989.36%90.60%+1.294.62%93.08%-1.557.75%59.90%+2.21 · 1942.0/1942.083.40%86.54%+3.189.18%90.51%+1.395.08%93.44%-1.657.43%59.69%+2.3
63 e2_dual_bot_phobert_v2_reply_gemma3_1b_pt_concat_ce_real_lora_m1m1 3 · 2358.0/2358.085.08%87.21% ±1.12+2.186.24%88.57% ±0.11+2.393.85%93.33% ±2.35-0.549.59%52.95% ±3.34+3.43 · 1942.0/1942.081.44%83.78% ±2.31+2.386.00%88.44% ±0.15+2.494.26%93.72% ±2.50-0.549.53%53.02% ±3.03+3.5
64 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_llrd09_lr3e5_m1m1 3 · 2358.0/2358.085.13%87.13% ±1.03+2.086.89%88.69% ±0.49+1.894.10%93.33% ±0.44-0.849.35%51.14% ±0.87+1.83 · 1942.0/1942.080.19%82.90% ±2.30+2.786.50%88.37% ±0.49+1.993.99%93.17% ±0.47-0.848.60%50.30% ±0.56+1.7
65 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr2e5_ep6_m1m1 3 · 2358.0/2358.085.03%86.62% ±1.96+1.686.90%88.14% ±1.56+1.292.05%90.26% ±2.70-1.851.77%51.57% ±5.09-0.23 · 1942.0/1942.080.26%82.22% ±3.77+2.086.54%87.91% ±1.58+1.492.35%90.44% ±2.88-1.950.93%50.85% ±5.09-0.1
66 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t2_tsa_m1m1 3 · 2358.0/2358.087.16%87.92% ±0.55+0.887.93%89.50% ±0.46+1.691.54%90.26% ±1.60-1.354.68%56.00% ±3.92+1.33 · 1942.0/1942.084.46%84.97% ±1.06+0.587.61%89.22% ±0.61+1.692.35%90.44% ±1.71-1.954.21%55.42% ±4.25+1.2
67 e2_dual_phobert_v2_seg_concat_ce_real_fullemb_emblr01_lr2e5_m1m1 3 · 2358.0/2358.085.50%87.25% ±1.21+1.887.20%89.07% ±0.80+1.994.10%93.33% ±0.89-0.853.44%54.45% ±4.56+1.03 · 1942.0/1942.081.04%82.91% ±2.44+1.986.84%88.75% ±0.82+1.994.26%93.44% ±0.82-0.852.69%53.64% ±4.72+1.0
68 e2_gemma3_1b_pt_cross_ce_real_lora_lr5e5_m1m1 3 · 2358.0/2358.081.15%83.17% ±4.68+2.087.39%89.02% ±1.32+1.690.77%88.21% ±0.44-2.653.06%56.43% ±5.94+3.43 · 1942.0/1942.077.58%80.31% ±2.96+2.787.10%88.87% ±1.39+1.890.71%88.25% ±0.47-2.553.09%56.67% ±6.46+3.6
69 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_kd_a05_t1_tsa_m1m1 3 · 2358.0/2358.084.79%86.98% ±0.37+2.286.68%88.92% ±0.54+2.294.10%91.79% ±1.18-2.349.81%53.55% ±3.86+3.73 · 1942.0/1942.080.85%83.28% ±0.97+2.486.32%88.70% ±0.52+2.495.08%92.62% ±0.82-2.549.08%52.89% ±3.59+3.8
70 e2_mmbert_base_cross_factor_ce_synthv2_frozenemb_lr2e5_conj_margin05_m1m1 3 · 2358.0/2358.083.72%85.26% ±1.02+1.586.09%88.36% ±0.15+2.393.85%92.56% ±2.35-1.347.34%51.45% ±0.81+4.13 · 1942.0/1942.079.65%81.05% ±3.57+1.485.99%88.41% ±0.14+2.494.26%92.90% ±2.50-1.447.88%52.34% ±1.19+4.5
71 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.084.72%86.91% ±1.10+2.287.64%90.02% ±0.55+2.492.56%91.03% ±1.78-1.553.52%58.03% ±1.21+4.53 · 1942.0/1942.077.00%79.33% ±1.09+2.387.35%89.91% ±0.62+2.693.17%91.80% ±2.17-1.453.00%57.76% ±1.41+4.8
72 e2_phobert_v2_seg_cross_ce_real_fullemb_lr2e5_m1m1 3 · 2358.0/2358.085.49%88.02% ±2.03+2.588.06%90.33% ±1.33+2.390.77%89.49% ±0.89-1.357.55%61.24% ±7.00+3.73 · 1942.0/1942.079.69%82.79% ±4.25+3.187.85%90.27% ±1.27+2.491.80%90.44% ±0.95-1.457.07%61.00% ±6.57+3.9
73 e2_dual_phobert_v2_seg_concat_mlp_ce_real_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.084.81%86.48% ±1.03+1.787.17%88.71% ±0.88+1.592.56%91.54% ±1.33-1.052.98%53.33% ±2.71+0.33 · 1942.0/1942.080.53%82.64% ±1.93+2.186.78%88.51% ±0.97+1.792.62%91.53% ±0.95-1.152.41%52.99% ±3.16+0.6
74 e2_xlmr_base_cross_logitadjust_real_frozenemb_m1m1 3 · 2358.0/2358.082.89%85.63% ±0.67+2.785.62%88.15% ±0.92+2.593.59%92.56% ±0.89-1.047.18%50.52% ±2.90+3.33 · 1942.0/1942.078.09%81.09% ±0.77+3.085.13%87.82% ±1.06+2.793.99%92.90% ±1.71-1.146.40%49.83% ±3.25+3.4
75 e2_me5_large_instruct_cross_ce_real_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.086.24%88.63% ±0.89+2.488.85%90.75% ±0.78+1.992.05%90.26% ±1.78-1.858.24%62.32% ±6.19+4.13 · 1942.0/1942.082.41%85.18% ±1.02+2.888.45%90.46% ±0.76+2.091.80%89.89% ±2.37-1.957.20%61.40% ±6.06+4.2
76 e2_mmbert_base_cross_factor_ce_real_frozenemb_lr2e5_conj_m1m1 3 · 2358.0/2358.084.02%85.79% ±1.12+1.886.95%88.94% ±0.35+2.089.23%87.44% ±4.70-1.854.26%58.10% ±3.91+3.83 · 1942.0/1942.079.33%81.12% ±2.25+1.886.53%88.63% ±0.40+2.189.34%87.43% ±5.01-1.953.28%57.23% ±3.66+3.9
77 e2_mmbert_base_cross_ce_real_frozenemb_lr3e5_m1m1 3 · 2358.0/2358.083.85%85.77% ±0.44+1.986.97%89.12% ±0.48+2.293.08%90.00% ±1.54-3.153.05%57.59% ±2.12+4.53 · 1942.0/1942.079.37%81.34% ±1.45+2.086.58%88.87% ±0.62+2.393.44%90.16% ±1.64-3.352.36%57.06% ±2.28+4.7
78 e2_dual_bot_phobert_v2_reply_qwen3_17b_base_concat_ce_real_lora_m1m1 1 · 2358.0/2358.083.94%85.75%+1.885.91%88.08%+2.293.08%93.08%+0.044.98%49.19%+4.21 · 1942.0/1942.080.17%82.11%+1.985.54%87.86%+2.393.44%93.44%+0.044.53%48.93%+4.4
79 e2_visobert_cross_ce_real_frozenemb_lr4e5_m1m1 1 · 2358.0/2358.081.50%82.96%+1.583.40%85.15%+1.793.08%92.31%-0.839.16%40.54%+1.41 · 1942.0/1942.077.56%79.13%+1.683.07%84.95%+1.993.44%92.62%-0.839.18%40.65%+1.5
80 e2_xlmr_base_cross_ce_real_frozenemb_lr4e5_m1m1 1 · 2358.0/2358.083.69%86.87%+3.286.02%88.59%+2.689.23%89.23%+0.047.35%51.33%+4.01 · 1942.0/1942.079.12%82.79%+3.785.72%88.44%+2.790.16%90.16%+0.047.21%51.40%+4.2
81 e2_dual_mmbert_concat_ce_real_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.085.18%86.53% ±0.93+1.387.47%89.09% ±1.39+1.693.59%91.79% ±1.60-1.850.41%53.60% ±6.21+3.23 · 1942.0/1942.079.32%80.79% ±2.08+1.587.22%88.98% ±1.25+1.893.72%92.08% ±1.71-1.650.15%53.57% ±5.27+3.4
82 e2_dual_phobert_v2_seg_concat_ce_synthm1blind_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.086.95%88.27% ±0.46+1.389.28%90.71% ±1.97+1.488.72%87.95% ±3.47-0.866.36%65.70% ±12.82-0.73 · 1942.0/1942.083.55%85.19% ±1.18+1.688.94%90.43% ±2.10+1.588.80%87.70% ±3.28-1.165.57%64.74% ±13.13-0.8
83 e2_dual_phobert_v2_seg_gated_ce_real_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.085.76%87.43% ±0.72+1.788.15%89.96% ±0.35+1.890.51%89.74% ±1.78-0.857.48%58.32% ±3.56+0.83 · 1942.0/1942.080.88%82.66% ±2.22+1.887.70%89.64% ±0.48+1.990.71%89.89% ±1.89-0.856.77%57.62% ±3.75+0.9
84 e2_dual_phobert_v2_seg_concat_ce_synth1x_frozenemb_lr2e5_fbmatched_m1m1 3 · 2358.0/2358.084.57%86.72% ±0.33+2.287.68%89.31% ±1.65+1.691.54%89.74% ±3.87-1.856.25%57.40% ±8.00+1.23 · 1942.0/1942.078.76%81.96% ±1.18+3.287.27%89.02% ±1.64+1.891.53%89.62% ±4.21-1.955.26%56.52% ±7.79+1.3
85 e2_dual_phobert_v2_seg_concat_ce_synth1x_fullemb_lr2e5_fbmatched_m1m1 3 · 2358.0/2358.084.59%85.74% ±0.40+1.287.37%88.95% ±1.34+1.692.56%90.51% ±2.70-2.155.37%55.59% ±5.54+0.23 · 1942.0/1942.078.83%80.05% ±1.85+1.287.09%88.81% ±1.29+1.792.62%90.44% ±2.88-2.255.01%55.41% ±5.35+0.4
86 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_hd01_m1m1 3 · 2358.0/2358.083.85%86.18% ±1.50+2.386.80%88.71% ±0.89+1.991.54%90.51% ±1.18-1.051.50%54.54% ±4.14+3.03 · 1942.0/1942.078.96%81.37% ±3.13+2.486.33%88.35% ±0.96+2.091.53%90.44% ±1.71-1.150.29%53.37% ±4.09+3.1
87 e2_mmbert_base_cross_ce_augasrnoise_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.084.98%86.95% ±0.66+2.086.76%88.87% ±1.06+2.191.03%89.23% ±4.62-1.851.91%54.54% ±5.99+2.63 · 1942.0/1942.082.24%84.13% ±1.55+1.986.38%88.60% ±1.05+2.290.98%89.07% ±4.51-1.951.25%53.98% ±5.72+2.7
88 e2_mmbert_base_cross_ce_real_frozenemb_lr2e5_pool4_rdrop1_msd5_m1m1 3 · 2358.0/2358.085.37%87.69% ±1.41+2.387.93%90.16% ±0.89+2.291.79%90.26% ±1.94-1.557.88%61.45% ±5.92+3.63 · 1942.0/1942.081.35%83.87% ±1.95+2.587.62%89.91% ±0.88+2.392.08%90.44% ±1.71-1.657.03%60.61% ±5.84+3.6
89 e2_qwen3_06b_base_genlabel_ce_real_lora_noeot_m1m1 3 · 2358.0/2358.085.05%87.36% ±1.75+2.387.47%89.58% ±2.23+2.193.59%93.59% ±1.18-0.050.42%55.62% ±8.24+5.23 · 1942.0/1942.081.55%84.41% ±2.42+2.987.05%89.33% ±2.35+2.393.44%93.44% ±1.42+0.049.73%55.22% ±8.43+5.5
90 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a09_t1_tsa_m1m1 3 · 2358.0/2358.086.54%87.42% ±0.82+0.987.53%89.13% ±0.59+1.693.59%91.79% ±2.35-1.853.51%55.01% ±3.36+1.53 · 1942.0/1942.083.41%83.75% ±2.48+0.387.16%88.86% ±0.68+1.793.72%91.80% ±2.17-1.952.86%54.40% ±3.64+1.5
91 e2_dual_bot_phobert_v2_reply_qwen3_06b_base_concat_ce_real_lora_m1m1 3 · 2358.0/2358.084.16%85.40% ±2.21+1.285.86%87.35% ±0.99+1.590.26%88.97% ±3.87-1.345.47%47.72% ±4.30+2.33 · 1942.0/1942.080.31%81.69% ±4.11+1.485.90%87.55% ±1.53+1.790.44%89.07% ±4.13-1.447.40%50.16% ±7.15+2.8
92 e2_dual_phobert_v2_seg_concat_ce_frozenemb_reinit2_lr3e5_synthrep_m1m1 3 · 2358.0/2358.086.17%87.99% ±0.91+1.887.95%89.57% ±0.44+1.692.56%91.28% ±1.78-1.355.82%56.78% ±0.15+1.03 · 1942.0/1942.082.81%84.89% ±1.45+2.187.57%89.26% ±0.52+1.792.62%91.26% ±2.37-1.454.88%55.85% ±0.16+1.0
93 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsphobert_m1m1 3 · 2358.0/2358.086.35%87.42% ±1.11+1.188.02%89.56% ±2.06+1.591.54%90.00% ±2.77-1.558.93%59.09% ±8.12+0.23 · 1942.0/1942.082.19%83.69% ±1.30+1.587.64%89.29% ±2.08+1.690.98%89.34% ±2.96-1.658.33%58.48% ±8.17+0.2
94 e2_xlmr_base_cross_ce_real_frozenemb_m1m1 3 · 2358.0/2358.083.20%85.94% ±1.86+2.786.03%88.28% ±1.37+2.290.26%90.26% ±3.55+0.050.47%54.11% ±8.05+3.63 · 1942.0/1942.077.52%80.40% ±3.98+2.985.71%88.04% ±1.57+2.390.98%90.98% ±3.57+0.049.96%53.67% ±8.66+3.7
95 e2_xlmr_base_cross_wcesqrt_real_frozenemb_m1m1 3 · 2358.0/2358.084.31%86.94% ±0.72+2.686.38%88.96% ±1.02+2.692.56%91.54% ±2.66-1.051.70%55.65% ±6.10+3.93 · 1942.0/1942.080.41%83.32% ±1.04+2.985.92%88.68% ±1.16+2.893.44%92.35% ±1.71-1.150.80%54.92% ±6.34+4.1
96 e2_dual_phobert_v2_seg_concat_logitadj025_real_frozenemb_lr2e5_m1m1 3 · 2358.0/2358.083.91%84.70% ±1.42+0.887.05%88.27% ±0.47+1.293.08%91.54% ±0.77-1.551.80%51.60% ±0.80-0.23 · 1942.0/1942.079.40%80.28% ±2.06+0.986.76%88.06% ±0.47+1.393.17%91.26% ±1.25-1.951.38%51.16% ±0.84-0.2
97 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr1e5_m1m1 3 · 2358.0/2358.083.15%84.00% ±4.29+0.987.48%88.88% ±0.88+1.491.03%90.00% ±2.04-1.056.24%55.57% ±7.22-0.73 · 1942.0/1942.080.71%81.30% ±2.29+0.687.22%88.66% ±0.85+1.491.53%89.34% ±2.17-2.255.95%55.16% ±6.84-0.8
98 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_reinit2_lr3e5_kd_a05_t2_tsa_m1m1 3 · 2358.0/2358.085.89%87.05% ±0.94+1.288.07%89.63% ±1.23+1.692.05%90.77% ±2.04-1.356.31%57.66% ±4.99+1.33 · 1942.0/1942.081.76%82.54% ±2.76+0.887.71%89.36% ±1.30+1.792.08%90.71% ±1.71-1.455.62%56.94% ±5.04+1.3
99 e2_qwen3emb4b_crossattn_wcesqrtls_real_frozen_m1m1 2 · 2358.0/2358.082.11%83.59% ±0.67+1.584.23%85.84% ±1.88+1.690.00%88.85% ±2.72-1.247.85%50.88% ±0.03+3.02 · 1942.0/1942.078.03%79.60% ±0.27+1.683.47%85.17% ±2.01+1.789.34%88.11% ±2.90-1.246.58%49.53% ±0.34+3.0
100 e2_phobert_v2_seg_cross_ce_real_frozenemb_lr1e5_m1m1 1 · 2358.0/2358.076.27%76.73%+0.586.53%89.22%+2.789.23%88.46%-0.854.98%57.50%+2.51 · 1942.0/1942.075.77%74.44%-1.386.47%89.32%+2.890.16%89.34%-0.855.84%58.60%+2.8
101 e2_qwen3emb06b_crossattn_wcesqrtls_real_frozen_m1m1 3 · 2358.0/2358.082.32%84.55% ±0.52+2.284.83%86.97% ±1.13+2.184.10%82.82% ±0.44-1.352.38%54.84% ±5.54+2.53 · 1942.0/1942.078.33%80.80% ±0.59+2.584.15%86.42% ±1.26+2.383.33%81.97% ±0.00-1.451.01%53.51% ±5.90+2.5
102 e2_dual_phobert_v2_seg_concat_ce_real_frozenemb_lr5e5_m1m1 3 · 2358.0/2358.085.20%86.62% ±0.35+1.487.48%88.72% ±0.95+1.293.59%92.05% ±1.94-1.553.58%54.09% ±5.37+0.53 · 1942.0/1942.080.89%82.56% ±0.15+1.787.07%88.41% ±0.98+1.393.72%92.08% ±2.06-1.652.68%53.18% ±5.32+0.5
103 e2_me5_base_cross_ce_real_frozenemb_lr1e5_m1m1 1 · 2358.0/2358.076.82%76.85%+0.085.51%87.51%+2.090.00%89.23%-0.847.95%50.88%+2.91 · 1942.0/1942.077.06%76.10%-1.085.20%87.32%+2.190.16%89.34%-0.848.03%51.17%+3.1
104 e2_visobert_cross_ce_real_frozenemb_lr2e5_m1m1 1 · 2358.0/2358.079.68%82.81%+3.183.95%85.46%+1.590.00%89.23%-0.841.79%43.28%+1.51 · 1942.0/1942.072.43%77.08%+4.683.58%85.22%+1.690.98%90.16%-0.841.57%43.14%+1.6
105 e2_visobert_cross_ce_real_frozenemb_lr1e5_m1m1 1 · 2358.0/2358.079.30%82.07%+2.882.85%83.94%+1.186.15%87.69%+1.544.09%45.42%+1.31 · 1942.0/1942.073.36%76.25%+2.982.31%83.50%+1.286.07%87.70%+1.643.57%44.96%+1.4
106 e2_xlmr_base_cross_ce_real_frozenemb_lr1e5_m1m1 1 · 2358.0/2358.080.10%84.03%+3.985.75%87.50%+1.890.00%90.00%+0.052.70%54.42%+1.71 · 1942.0/1942.073.36%79.91%+6.685.37%87.32%+2.090.98%90.98%+0.052.11%53.88%+1.8
107 e1_tfidf_hybrid_qwen3emb_lr_m1m1 1 · 2358.0/2358.081.01%82.08%+1.183.30%85.16%+1.984.62%83.85%-0.846.81%49.77%+3.01 · 1942.0/1942.075.09%74.39%-0.782.84%84.82%+2.084.43%83.61%-0.846.61%49.76%+3.1
108 e1_tfidf_svmcal_ce_synth1x_fbmatched_m1m1 1 · 2358.0/2358.083.64%84.20%+0.685.82%87.13%+1.381.54%79.23%-2.357.61%59.54%+1.91 · 1942.0/1942.077.77%78.33%+0.685.33%86.72%+1.481.15%78.69%-2.556.90%58.90%+2.0
109 e1_tfidf_svmcal_ce_synth1x_nofb_m1m1 1 · 2358.0/2358.082.76%83.43%+0.785.40%86.78%+1.480.77%78.46%-2.354.97%57.63%+2.71 · 1942.0/1942.076.80%77.51%+0.784.89%86.35%+1.580.33%77.87%-2.554.14%56.89%+2.7
110 e1_tfidf_svmcal_ce_synth1x_noqparticle_m1m1 1 · 2358.0/2358.083.27%83.13%-0.183.92%85.35%+1.484.62%83.85%-0.846.81%48.88%+2.11 · 1942.0/1942.079.82%78.52%-1.383.48%85.00%+1.584.43%83.61%-0.846.61%48.80%+2.2
112 e1_tfidf_svmcal_ce_synth1x_nofiller_m1m1 1 · 2358.0/2358.082.99%84.50%+1.583.93%85.74%+1.887.69%86.92%-0.846.91%49.34%+2.41 · 1942.0/1942.078.49%80.08%+1.683.28%85.19%+1.987.70%86.89%-0.845.73%48.18%+2.5
113 e1_tfidf_svmcal_ce_synth1x_noofftopic_m1m1 1 · 2358.0/2358.083.02%82.92%-0.184.23%85.65%+1.484.62%83.85%-0.848.25%50.23%+2.01 · 1942.0/1942.079.44%78.32%-1.183.90%85.42%+1.584.43%83.61%-0.848.58%50.75%+2.2
114 e2_dual_qwen3emb06b_concat_ce_real_frozen_m1m1 3 · 2358.0/2358.080.49%82.22% ±1.31+1.781.14%83.74% ±1.22+2.685.90%86.15% ±0.00+0.345.40%48.94% ±4.87+3.53 · 1942.0/1942.074.43%76.77% ±1.64+2.380.44%83.16% ±1.21+2.785.52%85.25% ±0.00-0.344.57%48.08% ±4.71+3.5
115 e1_tfidf_svmcal_ce_synth1x_base_m1m1 1 · 2358.0/2358.082.56%82.62%+0.183.68%85.30%+1.683.85%83.08%-0.846.58%48.65%+2.11 · 1942.0/1942.078.84%77.89%-1.083.18%84.89%+1.783.61%82.79%-0.846.15%48.33%+2.2
116 e1_tfidf_svmcal_ce_synth1x_matched_m1m1 1 · 2358.0/2358.082.76%83.20%+0.485.40%86.51%+1.183.08%80.77%-2.355.10%56.76%+1.71 · 1942.0/1942.076.78%77.26%+0.584.86%86.04%+1.282.79%80.33%-2.554.30%56.00%+1.7
117 e1_tfidf_svmcal_ce_synth1x_nomoney_m1m1 1 · 2358.0/2358.083.74%83.26%-0.584.49%85.51%+1.085.38%83.85%-1.548.68%50.46%+1.81 · 1942.0/1942.080.24%78.59%-1.683.98%85.08%+1.185.25%83.61%-1.648.15%50.00%+1.9
118 e1_tfidf_svmcal_ce_synth1x_fbonly_m1m1 1 · 2358.0/2358.082.32%81.80%-0.583.40%84.32%+0.985.38%83.85%-1.543.87%43.78%-0.11 · 1942.0/1942.078.78%77.22%-1.683.11%84.10%+1.085.25%83.61%-1.644.64%44.54%-0.1

Phân tích theo mức ảnh hưởng

MứcMethodRunKết luận
highe2_xlmr_base_cross_ce_real_frozenembe2_xlmr_base_cross_ce_real_frozenemb_seed13_0a1472e9On 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