List your company

Backdoor Mitigation on SFT-based Poisoning Phrase trigger

95.7Clean Accuracy (CACC)SFT
92.028892.981993.93594.8881Oct 11, 2025

Evaluation results

MethodLinks
SFTBackbone=Llama-3-8B2025.1095.7100
BaseBackbone=Llama-3-8B2025.1095.6100
LocphylaxBackbone=Llama-3-8B2025.1095.40
Fine-MixingBackbone=Qwen-2.5-7B2025.10951.25
SFTBackbone=Qwen-2.5-7B2025.109584.73
LocphylaxBackbone=Mistral-7B2025.1094.939.2
DistillationBackbone=Mistral-7B2025.1094.993.2
PDBBackbone=Qwen-2.5-7B2025.1094.959
DistillationBackbone=Qwen-2.5-7B2025.1094.878
LocphylaxBackbone=Qwen-2.5-7B2025.1094.60
Fine-MixingBackbone=Mistral-7B2025.1094.2249.2
DistillationBackbone=Llama-3-8B2025.109464
PDBBackbone=Llama-3-8B2025.1093.94
BaseBackbone=Qwen-2.5-7B2025.1093.4100
Fine-MixingBackbone=Llama-3-8B2025.1093.312.75
SFTBackbone=Mistral-7B2025.1093.122.8
PDBBackbone=Mistral-7B2025.1092.516
BaseBackbone=Mistral-7B2025.1092.1796.6