Natural Language Inference on MNLI (Acc, BT, DR, ASR)
85.27AccuracySafeLoRA
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| SafeLoRABackbone=Qwen2.5-1.5B2026.06 | 85.27 | 49.51 | 45.79 | 47.65 | |
| SafeMERGEBackbone=Qwen2.5-1.5B2026.06 | 84.53 | 40.28 | 36.58 | 38.43 | |
| SafeGeneBackbone=Qwen2.5-1.5B2026.06 | 84.45 | 32.49 | 30 | 31.25 | |
| Fine-tunedBackbone=Qwen2.5-1.5B2026.06 | 84.23 | 51.93 | 48.82 | 50.38 | |
| SaLoRABackbone=Qwen2.5-1.5B2026.06 | 84.17 | 48.36 | 45.13 | 46.75 |