Question Answering on BoolQ (Accuracy, BT, DR, ASR)
82.34AccuracySafeLoRA
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SafeLoRABackbone=Qwen2.5-1.5B2026.06 | 82.34 | 49.91 | 44.61 | 47.26 | |
| SafeMERGEBackbone=Qwen2.5-1.5B2026.06 | 81.04 | 42.99 | 36.18 | 39.59 | |
| Fine-tunedBackbone=Qwen2.5-1.5B2026.06 | 80.28 | 52.28 | 45.66 | 48.97 | |
| SafeGeneBackbone=Qwen2.5-1.5B2026.06 | 79.63 | 31.62 | 27.63 | 29.63 | |
| SaLoRABackbone=Qwen2.5-1.5B2026.06 | 78.7 | 44.2 | 43.95 | 44.08 |