List your company

Safety-constrained Fine-tuning on ARC Easy

97UtilityNo defense
86.689.39294.7May 13, 2026

Evaluation results

MethodLinks
No defenseDefense=No defense2026.0597964.91
SafeLoRADefense=SafeLoRA2026.0597964.96
ModerationDefense=Moderation2026.0596984.7
SafeInstrDefense=SafeInstr2026.0595944.94
BackdoorDefense=Backdoor2026.0595613.42
GradShieldDefense=GradShield2026.059441.13
baseDefense=base2026.058741.16