Safety-constrained Fine-tuning on GSM8k
87UtilityGradShield
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GradShieldDefense=GradShield2026.05 | 87 | 0 | 1 | |
| baseDefense=base2026.05 | 86 | 4 | 1.16 | |
| ModerationDefense=Moderation2026.05 | 85 | 52 | 2.4 | |
| BackdoorDefense=Backdoor2026.05 | 84 | 78 | 4.22 | |
| No defenseDefense=No defense2026.05 | 83 | 97 | 5 | |
| SafeInstrDefense=SafeInstr2026.05 | 83 | 96 | 4.98 | |
| SafeLoRADefense=SafeLoRA2026.05 | 81 | 98 | 4.99 |