Backdoor Mitigation on SFT-based Poisoning Phrase trigger
95.7Clean Accuracy (CACC)SFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFTBackbone=Llama-3-8B2025.10 | 95.7 | 100 | |
| BaseBackbone=Llama-3-8B2025.10 | 95.6 | 100 | |
| LocphylaxBackbone=Llama-3-8B2025.10 | 95.4 | 0 | |
| Fine-MixingBackbone=Qwen-2.5-7B2025.10 | 95 | 1.25 | |
| SFTBackbone=Qwen-2.5-7B2025.10 | 95 | 84.73 | |
| LocphylaxBackbone=Mistral-7B2025.10 | 94.93 | 9.2 | |
| DistillationBackbone=Mistral-7B2025.10 | 94.9 | 93.2 | |
| PDBBackbone=Qwen-2.5-7B2025.10 | 94.9 | 59 | |
| DistillationBackbone=Qwen-2.5-7B2025.10 | 94.8 | 78 | |
| LocphylaxBackbone=Qwen-2.5-7B2025.10 | 94.6 | 0 | |
| Fine-MixingBackbone=Mistral-7B2025.10 | 94.22 | 49.2 | |
| DistillationBackbone=Llama-3-8B2025.10 | 94 | 64 | |
| PDBBackbone=Llama-3-8B2025.10 | 93.9 | 4 | |
| BaseBackbone=Qwen-2.5-7B2025.10 | 93.4 | 100 | |
| Fine-MixingBackbone=Llama-3-8B2025.10 | 93.3 | 12.75 | |
| SFTBackbone=Mistral-7B2025.10 | 93.1 | 22.8 | |
| PDBBackbone=Mistral-7B2025.10 | 92.5 | 16 | |
| BaseBackbone=Mistral-7B2025.10 | 92.17 | 96.6 |