Text Classification on QNLI
92.79Accuracy (%)theta_B fine-tune
Evaluation Results
| Method | Links | |
|---|---|---|
| theta_B fine-tune|Ds^c|=-2025.10 | 92.79 | |
| theta_B + delta*|Ds^c|=-2025.10 | 65.32 | |
| Angular (γ=8)gamma=8, Sequence Length=20482025.10 | 61.7 | |
| RACE (P=4, L=4)P=4, L=4, Sequence Length=20482025.10 | 61.1 | |
| FlashAttention2Sequence Length=20482025.10 | 61.1 | |
| SigmoidSequence Length=20482025.10 | 61.1 | |
| Performer-256Dimension=256, Sequence Length=20482025.10 | 61 | |
| RACE (P=2, L=2)P=2, L=2, Sequence Length=20482025.10 | 60.7 | |
| RACE (P=3, L=3)P=3, L=3, Sequence Length=20482025.10 | 60.7 | |
| LinearSequence Length=20482025.10 | 60.7 | |
| Linformer-128Projection Dimension=128, Sequence Length=20482025.10 | 60.6 | |
| GradFix (theta_B + delta^A)|Ds^c|=502025.10 | 60.5 | |
| theta_B^opt|Ds^c|=502025.10 | 51.45 | |
| theta_B zero-shot|Ds^c|=-2025.10 | 50.54 | |
| theta_B + tau_A|Ds^c|=-2025.10 | 50.52 |