List your company

Backdoor Mitigation on SFT-based Poisoning Word trigger

96.7Clean Accuracy (CACC)SFT
86.528889.169491.8194.4506Oct 11, 2025

Evaluation results

MethodLinks
SFTBackbone=Llama-3-8B2025.1096.778.61
LocphylaxBackbone=Llama-3-8B2025.1096.27.5
DistillationBackbone=Mistral-7B2025.1095.80.98
SFTBackbone=Mistral-7B2025.1095.71.6
BaseBackbone=Mistral-7B2025.1095.590.2
BaseBackbone=Llama-3-8B2025.1095.188.44
LocphylaxBackbone=Mistral-7B2025.1095.030.64
Fine-MixingBackbone=Mistral-7B2025.1094.932.8
BaseBackbone=Qwen-2.5-7B2025.1094.698
Fine-MixingBackbone=Llama-3-8B2025.1094.617.22
Fine-MixingBackbone=Qwen-2.5-7B2025.1094.68.25
SFTBackbone=Qwen-2.5-7B2025.1094.686
DistillationBackbone=Llama-3-8B2025.1094.254
LocphylaxBackbone=Qwen-2.5-7B2025.1094.28.25
PDBBackbone=Llama-3-8B2025.1094.16.39
PDBBackbone=Qwen-2.5-7B2025.109430
DistillationBackbone=Qwen-2.5-7B2025.1093.898
PDBBackbone=Mistral-7B2025.1086.923.6