Natural Language Inference on NLI domain average
87.5Best AccuracyTask-level fine-tuning
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Task-level fine-tuning2025.06 | 87.5 | — | |
| Many-shot (ours)Masking Strategy=mask all targets2025.06 | 86.8 | 86.5 | |
| Many-shot (ours)Masking Strategy=mask last target2025.06 | 85.5 | 84.9 | |
| Few-shotMasking Strategy=mask all targets2025.06 | 83.9 | 83.2 | |
| Few-shotMasking Strategy=mask last target2025.06 | 83.2 | 82.4 | |
| Many-shotStrategy=autoregressive2025.06 | 81.7 | 81.7 | |
| Zero shot2025.06 | 81.4 | 80.6 | |
| No fine-tuning2025.06 | 78.9 | 74.7 |