List your company

Safety-constrained Fine-tuning on AGNews

92UtilityNo defense
74.3278.9183.588.09May 13, 2026

Evaluation results

MethodLinks
No defenseDefense=No defense2026.0592994.98
SafeInstrDefense=SafeInstr2026.0592934.83
GradShieldDefense=GradShield2026.059151.19
ModerationDefense=Moderation2026.0590874.45
SafeLoRADefense=SafeLoRA2026.0590944.94
BackdoorDefense=Backdoor2026.0588774.15
baseDefense=base2026.057541.16