List your company

Harmful Fine-tuning Attack on BeaverTails

64.5Attack Success Rate (ASR)SFT
6.57221.61136.6551.689Jun 29, 2026

Evaluation results

MethodLinks
SFTModel=Gemma-2-9B-IT2026.0664.5
SFTModel=Qwen2.5-7B-Instruct2026.0663.6
SFTModel=Llama-2-7B-Chat2026.0662.5
LisaModel=Llama-2-7B-Chat2026.0661.4
LisaModel=Gemma-2-9B-IT2026.0660.8
LisaModel=Qwen2.5-7B-Instruct2026.0660.1
Constrained SFTModel=Gemma-2-9B-IT2026.0647.7
Constrained SFTModel=Qwen2.5-7B-Instruct2026.0646.4
Constrained SFTModel=Llama-2-7B-Chat2026.0637.8
DR-SFTModel=Qwen2.5-7B-Instruct2026.0617
DR-SFTModel=Llama-2-7B-Chat2026.0614.3
DR-SFTModel=Gemma-2-9B-IT2026.068.8