List your company

Malicious Fine-tuning Attack on HEx-PHI

95.1Attack Success Rate (ASR)SFT
-1.93223.25948.4573.641Jun 6, 2026Jun 9, 2026Jun 13, 2026Jun 17, 2026Jun 21, 2026Jun 25, 2026Jun 29, 2026

Evaluation results

MethodLinks
SFTModel=Qwen2.5-7B-Instruct2026.0695.1
SFTModel=Gemma-2-9B-IT2026.0691.5
SFTModel=Llama-2-7B-Chat2026.0687.8
LisaModel=Qwen2.5-7B-Instruct2026.0681.3
VaccineTest-time Attack Dataset=PKU-SafeRLHF-unsafe2026.0672.3
VaccineTest-time Attack Dataset=Beavertails-unsafe2026.0671.7
SFTTest-time Attack Dataset=Toxic-DPO2026.0671.1
VaccineTest-time Attack Dataset=Toxic-DPO2026.0671.1
BoosterTest-time Attack Dataset=Toxic-DPO2026.0667.9
SFTTest-time Attack Dataset=PKU-SafeRLHF-unsafe2026.0666.6
BoosterTest-time Attack Dataset=PKU-SafeRLHF-unsafe2026.0666.5
BoosterTest-time Attack Dataset=Beavertails-unsafe2026.0663.9
LisaModel=Llama-2-7B-Chat2026.0660.9
SFTTest-time Attack Dataset=Beavertails-unsafe2026.0658.6
PatcherTest-time Attack Dataset=Toxic-DPO2026.0657.3
PatcherTest-time Attack Dataset=PKU-SafeRLHF-unsafe2026.0647.8
LisaModel=Gemma-2-9B-IT2026.0628.7
PatcherTest-time Attack Dataset=Beavertails-unsafe2026.0626.5
Constrained SFTModel=Llama-2-7B-Chat2026.0625.2
Constrained SFTModel=Qwen2.5-7B-Instruct2026.0621.5
Constrained SFTModel=Gemma-2-9B-IT2026.0610.1
DR-SFTModel=Gemma-2-9B-IT2026.063.9
DR-SFTModel=Llama-2-7B-Chat2026.062.1
DR-SFTModel=Qwen2.5-7B-Instruct2026.061.8