Safety-constrained Fine-tuning on AGNews
92UtilityNo defense
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| No defenseDefense=No defense2026.05 | 92 | 99 | 4.98 | |
| SafeInstrDefense=SafeInstr2026.05 | 92 | 93 | 4.83 | |
| GradShieldDefense=GradShield2026.05 | 91 | 5 | 1.19 | |
| ModerationDefense=Moderation2026.05 | 90 | 87 | 4.45 | |
| SafeLoRADefense=SafeLoRA2026.05 | 90 | 94 | 4.94 | |
| BackdoorDefense=Backdoor2026.05 | 88 | 77 | 4.15 | |
| baseDefense=base2026.05 | 75 | 4 | 1.16 |