Text Generation on VPI Generation Tasks Llama3-8B Mistral-7B (test)
100ASRBackdoor
Evaluation results
| Method | Links | ||
|---|---|---|---|
| BackdoorBackbone=Mistral-7B2026.02 | 100 | 99 | |
| BackdoorBackbone=Llama3-8B2026.02 | 99 | 97.5 | |
| NADBackbone=Mistral-7B2026.02 | 90.2 | 93.5 | |
| WANBackbone=Mistral-7B2026.02 | 56 | 73.5 | |
| LETBackbone=Mistral-7B2026.02 | 32.5 | 88.5 | |
| BEEBackbone=Mistral-7B2026.02 | 28 | 92 | |
| CROWBackbone=Mistral-7B2026.02 | 25.5 | 78.5 | |
| WANBackbone=Llama3-8B2026.02 | 23 | 27.5 | |
| F/TBackbone=Llama3-8B2026.02 | 19 | 79.5 | |
| CROWBackbone=Llama3-8B2026.02 | 15 | 82 | |
| LETBackbone=Llama3-8B2026.02 | 12 | 87 | |
| NADBackbone=Llama3-8B2026.02 | 10.5 | 89 | |
| BEEBackbone=Llama3-8B2026.02 | 9.5 | 92.5 | |
| F/TBackbone=Mistral-7B2026.02 | 9.2 | 75 | |
| PROTOPURIFYBackbone=Mistral-7B2026.02 | 9 | 95 | |
| PROTOPURIFYBackbone=Llama3-8B2026.02 | 8 | 94.5 |