Backdoor Mitigation on SFT-based Poisoning Long trigger
94.8Clean Accuracy (CACC)Base
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BaseBackbone=Qwen-2.5-7B2025.10 | 94.8 | 100 | |
| SFTBackbone=Qwen-2.5-7B2025.10 | 94.8 | 98 | |
| LocphylaxBackbone=Llama-3-8B2025.10 | 94.7 | 1.25 | |
| Fine-MixingBackbone=Mistral-7B2025.10 | 94.62 | 4.2 | |
| BaseBackbone=Llama-3-8B2025.10 | 94.6 | 94.5 | |
| DistillationBackbone=Qwen-2.5-7B2025.10 | 94.6 | 92 | |
| PDBBackbone=Llama-3-8B2025.10 | 94.6 | 44 | |
| PDBBackbone=Qwen-2.5-7B2025.10 | 94.6 | 28 | |
| SFTBackbone=Llama-3-8B2025.10 | 94.6 | 92.25 | |
| LocphylaxBackbone=Qwen-2.5-7B2025.10 | 94.34 | 5 | |
| Fine-MixingBackbone=Llama-3-8B2025.10 | 94 | 52.5 | |
| DistillationBackbone=Llama-3-8B2025.10 | 93.8 | 48 | |
| LocphylaxBackbone=Mistral-7B2025.10 | 93.8 | 7.8 | |
| DistillationBackbone=Mistral-7B2025.10 | 93.6 | 86 | |
| Fine-MixingBackbone=Qwen-2.5-7B2025.10 | 93.2 | 18.75 | |
| BaseBackbone=Mistral-7B2025.10 | 92.68 | 100 | |
| SFTBackbone=Mistral-7B2025.10 | 90.5 | 100 | |
| PDBBackbone=Mistral-7B2025.10 | 89.55 | 100 |