Natural Language Inference (NLI) (test)
81.17AccuracyNEZHA-wwm (L)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| NEZHA-wwm (L)Whole Word Masking=true, Backbone Scale=Large2019.11 | 81.17 | — | |
| ERNIE 2.0 (L)Backbone Scale=Large2019.11 | 81 | — | |
| NEZHA (L)Backbone Scale=Large2019.11 | 80.44 | — | |
| ERNIE 2.0 (B)Backbone Scale=Base2019.11 | 79.7 | — | |
| NEZHA (B)Backbone Scale=Base2019.11 | 79.32 | — | |
| ZEN (P)Initialization=Pre-trained, Backbone Scale=Base2019.11 | 79.2 | — | |
| NEZHA-WWM (B)Whole Word Masking=true, Backbone Scale=Base2019.11 | 79.11 | — | |
| ERNIE 1.0Backbone Scale=Base2019.11 | 78.4 | — | |
| BERT-WWMWhole Word Masking=true, Backbone Scale=Base2019.11 | 78 | — | |
| BERT (P)Initialization=Pre-trained, Backbone Scale=Base2019.11 | 77.52 | — | |
| ZEN (R)Initialization=Random, Backbone Scale=Base2019.11 | 77.03 | — | |
| SkylineBackbone=multilingual e5, Target Language Adaptation=Korean2026.04 | 75.93 | — | |
| BERT (R)Initialization=Random, Backbone Scale=Base2019.11 | 75.01 | — | |
| XITE-reg-regBackbone=multilingual e5, Target Language Adaptation=Korean2026.04 | 57.56 | — | |
| XITE-reg-ldaBackbone=multilingual e5, Target Language Adaptation=Korean2026.04 | 54.79 | — | |
| SkylineBackbone=multilingual e5, Target Language Adaptation=Arabic2026.04 | 53.25 | — | |
| RA-CAShots=250, Selection Strategy=Checkpoint averaging, Run Type=Ensemble2023.05 | 50.5 | 0.3 | |
| RA-CAShots=50, Selection Strategy=Checkpoint averaging, Run Type=Ensemble2023.05 | 45 | 1.6 | |
| RA-LASTShots=50, Selection Strategy=Last checkpoint, Run Type=Ensemble2023.05 | 44.6 | 2.1 | |
| CAShots=50, Selection Strategy=Checkpoint averaging, Run Type=Single Run2023.05 | 44.4 | 1.9 | |
| XITE-reg-ldaBackbone=multilingual e5, Target Language Adaptation=Arabic2026.04 | 38.84 | — | |
| TRG-DEVShots=5, Selection Strategy=Early stopping (target dev), Run Type=Single Run2023.05 | 38.3 | 1.8 | |
| RA-CAShots=5, Selection Strategy=Checkpoint averaging, Run Type=Ensemble2023.05 | 38.3 | 1.1 | |
| RA-LASTShots=5, Selection Strategy=Last checkpoint, Run Type=Ensemble2023.05 | 38.2 | 1 | |
| XITE-reg-ldaBackbone=multilingual e5, Target Language Adaptation=Hindi2026.04 | 38.06 | — | |
| CAShots=5, Selection Strategy=Checkpoint averaging, Run Type=Single Run2023.05 | 37.6 | 1.4 | |
| GS-LASTShots=5, Selection Strategy=Group selection last, Run Type=Single Run2023.05 | 37.5 | 1.8 | |
| LASTShots=5, Selection Strategy=Final checkpoint, Run Type=Single Run2023.05 | 37 | 1.3 | |
| SRC-DEVShots=5, Selection Strategy=Early stopping (source dev), Run Type=Single Run2023.05 | 36.9 | 1.3 | |
| XITE-reg-regBackbone=multilingual e5, Target Language Adaptation=Arabic2026.04 | 34.35 | — | |
| XITE-reg-ldaBackbone=multilingual e5, Target Language Adaptation=Urdu2026.04 | 33.81 | — | |
| XITE-reg-regBackbone=multilingual e5, Target Language Adaptation=Hindi2026.04 | 27.42 | — | |
| XITE-reg-regBackbone=multilingual e5, Target Language Adaptation=Urdu2026.04 | 23.83 | — | |
| SkylineBackbone=multilingual e5, Target Language Adaptation=Hindi2026.04 | 18.5 | — | |
| Baseline-PSBackbone=multilingual e5, Target Language Adaptation=Korean2026.04 | 15.57 | — | |
| SkylineBackbone=multilingual e5, Target Language Adaptation=Urdu2026.04 | 5.63 | — | |
| Baseline-PSBackbone=multilingual e5, Target Language Adaptation=Urdu2026.04 | 0.66 | — | |
| Baseline-PSBackbone=multilingual e5, Target Language Adaptation=Hindi2026.04 | 0.1 | — | |
| Baseline-PSBackbone=multilingual e5, Target Language Adaptation=Arabic2026.04 | 0.04 | — |