Recognizing Textual Entailment on RTE (Adversarial Robustness)
100Repair AccuracyAll-layers Full FT
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| All-layers Full FTModel=DistilBERT2026.04 | 100 | 100 | 80.8 | 36.3 | |
| WARPModel=DistilBERT2026.04 | 100 | 100 | 83.8 | 44.1 | |
| All-layers Full FTModel=BERT2026.04 | 100 | 98 | 87.4 | 40.1 | |
| WARPModel=BERT2026.04 | 100 | 100 | 86.2 | 47.9 | |
| Same-layer LoRAModel=DistilBERT, rank=22026.04 | 62.7 | 97.5 | 97 | 32.3 | |
| Same-layer LoRAModel=BERT, rank=22026.04 | 45.5 | 98 | 99 | 40.6 | |
| OriginalModel=BERT2026.04 | 10 | 97.5 | 99 | 39.3 | |
| OriginalModel=DistilBERT2026.04 | 0 | 100 | 99.2 | 37.4 |