List your company

Safety-constrained Fine-tuning on GSM8k

87UtilityGradShield
80.7682.388485.62May 13, 2026

Evaluation results

MethodLinks
GradShieldDefense=GradShield2026.058701
baseDefense=base2026.058641.16
ModerationDefense=Moderation2026.0585522.4
BackdoorDefense=Backdoor2026.0584784.22
No defenseDefense=No defense2026.0583975
SafeInstrDefense=SafeInstr2026.0583964.98
SafeLoRADefense=SafeLoRA2026.0581984.99