Natural Language Inference on GLUE and SNLI v1 (test val)
91.22MNLI AccuracyθA fine-tune
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| θA fine-tuneN (Calibration Samples)=All, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 91.22 | 95.32 | 84.83 | 91.05 | 91.29 | 90.74 | — | |
| θB fine-tuneN (Calibration Samples)=All, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 85.28 | 93.59 | 77.25 | 89.57 | 88.51 | 86.84 | — | |
| BiCoN (Calibration Samples)=200, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 71.06 | 85.74 | 68.95 | 88.25 | 66.59 | 76.11 | 5.01 | |
| THESEUSN (Calibration Samples)=200, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 69.92 | 83.63 | 68.23 | 88.09 | 64.31 | 74.83 | 3.73 | |
| θAN (Calibration Samples)=200, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 62.49 | 81.18 | 65.7 | 86.23 | 65.94 | 72.31 | 1.21 | |
| BiCoN (Calibration Samples)=100, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 61.94 | 79.33 | 67.5 | 84.31 | 58.43 | 70.3 | 8 | |
| THESEUSN (Calibration Samples)=100, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 60.93 | 78.34 | 66.78 | 63.19 | 56.38 | 65.12 | 2.82 | |
| θBN (Calibration Samples)=200, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 60.83 | 82.45 | 63.54 | 85.52 | 63.18 | 71.1 | 0 | |
| θAN (Calibration Samples)=100, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 60.13 | 78.46 | 59.93 | 68.12 | 58.32 | 64.99 | 2.69 | |
| θBN (Calibration Samples)=100, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 58.48 | 75.43 | 61.37 | 62.44 | 53.78 | 62.3 | 0 | |
| BiCoN (Calibration Samples)=50, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 46.32 | 76.71 | 67.87 | 83.4 | 60.8 | 67.02 | 12.32 | |
| THESEUSN (Calibration Samples)=50, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 46.12 | 72.65 | 66.42 | 63.18 | 54.61 | 60.59 | 5.89 | |
| θAN (Calibration Samples)=50, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 39.34 | 75.58 | 58.12 | 68.62 | 53.21 | 58.97 | 4.27 | |
| θBN (Calibration Samples)=50, Source Model=T5-3B, Target Model=T5-Large, Protocol=linear probing2026.05 | 38.74 | 69.5 | 57.4 | 62.56 | 45.28 | 54.7 | 0 |