Backdoor Mitigation on SFT-based Poisoning Word trigger
96.7Clean Accuracy (CACC)SFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFTBackbone=Llama-3-8B2025.10 | 96.7 | 78.61 | |
| LocphylaxBackbone=Llama-3-8B2025.10 | 96.2 | 7.5 | |
| DistillationBackbone=Mistral-7B2025.10 | 95.8 | 0.98 | |
| SFTBackbone=Mistral-7B2025.10 | 95.7 | 1.6 | |
| BaseBackbone=Mistral-7B2025.10 | 95.5 | 90.2 | |
| BaseBackbone=Llama-3-8B2025.10 | 95.1 | 88.44 | |
| LocphylaxBackbone=Mistral-7B2025.10 | 95.03 | 0.64 | |
| Fine-MixingBackbone=Mistral-7B2025.10 | 94.93 | 2.8 | |
| BaseBackbone=Qwen-2.5-7B2025.10 | 94.6 | 98 | |
| Fine-MixingBackbone=Llama-3-8B2025.10 | 94.6 | 17.22 | |
| Fine-MixingBackbone=Qwen-2.5-7B2025.10 | 94.6 | 8.25 | |
| SFTBackbone=Qwen-2.5-7B2025.10 | 94.6 | 86 | |
| DistillationBackbone=Llama-3-8B2025.10 | 94.2 | 54 | |
| LocphylaxBackbone=Qwen-2.5-7B2025.10 | 94.2 | 8.25 | |
| PDBBackbone=Llama-3-8B2025.10 | 94.1 | 6.39 | |
| PDBBackbone=Qwen-2.5-7B2025.10 | 94 | 30 | |
| DistillationBackbone=Qwen-2.5-7B2025.10 | 93.8 | 98 | |
| PDBBackbone=Mistral-7B2025.10 | 86.92 | 3.6 |