Safety-constrained Fine-tuning on ARC Easy
97UtilityNo defense
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| No defenseDefense=No defense2026.05 | 97 | 96 | 4.91 | |
| SafeLoRADefense=SafeLoRA2026.05 | 97 | 96 | 4.96 | |
| ModerationDefense=Moderation2026.05 | 96 | 98 | 4.7 | |
| SafeInstrDefense=SafeInstr2026.05 | 95 | 94 | 4.94 | |
| BackdoorDefense=Backdoor2026.05 | 95 | 61 | 3.42 | |
| GradShieldDefense=GradShield2026.05 | 94 | 4 | 1.13 | |
| baseDefense=base2026.05 | 87 | 4 | 1.16 |