List your company

Mathematical Reasoning on GSM8K (Acc, SR, JB, HB)

81.65AccuracySFT
54.92261.86168.875.739Mar 8, 2026

Evaluation results

MethodLinks
SFTBackbone=Qwen2.5-7B-Instruct, Harmful data proportion (P)=10.0%2026.0381.6571.2552.594.5
PACTBackbone=Qwen2.5-7B-Instruct, Harmful data proportion (P)=10.0%2026.0380.899.271129.5
AsFTBackbone=Qwen2.5-7B-Instruct, Harmful data proportion (P)=10.0%2026.0380.59.272848
Safe LoRABackbone=Qwen2.5-7B-Instruct, Harmful data proportion (P)=10.0%2026.0373.6915.652044.5
InitialBackbone=Qwen2.5-7B-Instruct, Harmful data proportion (P)=10.0%2026.0356.561.637
Constrained SFTBackbone=Qwen2.5-7B-Instruct, Harmful data proportion (P)=10.0%2026.0355.9510.9230.580