Safety-constrained Fine-tuning on ARC Challenge
82UtilityModeration
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ModerationDefense=Moderation2026.05 | 82 | 91 | 4.42 | |
| SafeInstrDefense=SafeInstr2026.05 | 82 | 95 | 4.87 | |
| SafeLoRADefense=SafeLoRA2026.05 | 82 | 99 | 4.95 | |
| No defenseDefense=No defense2026.05 | 81 | 99 | 4.96 | |
| BackdoorDefense=Backdoor2026.05 | 80 | 75 | 4.12 | |
| GradShieldDefense=GradShield2026.05 | 79 | 5 | 1.16 | |
| baseDefense=base2026.05 | 74 | 4 | 1.16 |