Text Generation on AutoPoison Generation Llama3-8B Mistral-7B (test)
82.7ASRBackdoor
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BackdoorBackbone=Mistral-7B2026.02 | 82.7 | 99.9 | |
| NADBackbone=Mistral-7B2026.02 | 81.4 | 90.6 | |
| BackdoorBackbone=Llama3-8B2026.02 | 80 | 99.9 | |
| WANBackbone=Mistral-7B2026.02 | 77.5 | 90.1 | |
| NADBackbone=Llama3-8B2026.02 | 61.8 | 90.5 | |
| CROWBackbone=Mistral-7B2026.02 | 26.6 | 83.5 | |
| CROWBackbone=Llama3-8B2026.02 | 22.7 | 80.6 | |
| F/TBackbone=Llama3-8B2026.02 | 21 | 90.8 | |
| LETBackbone=Llama3-8B2026.02 | 14.3 | 92.1 | |
| BEEBackbone=Mistral-7B2026.02 | 11 | 96 | |
| F/TBackbone=Mistral-7B2026.02 | 8.3 | 82.4 | |
| PROTOPURIFYBackbone=Mistral-7B2026.02 | 8 | 96 | |
| WANBackbone=Llama3-8B2026.02 | 6.7 | 57.2 | |
| PROTOPURIFYBackbone=Llama3-8B2026.02 | 6.5 | 95.5 | |
| LETBackbone=Mistral-7B2026.02 | 6.2 | 89 | |
| BEEBackbone=Llama3-8B2026.02 | 5 | 95 |