Safety Robustness on Prefill Attack
3.8ASRPrefill
Evaluation Results
| Method | Links | |
|---|---|---|
| PrefillBackbone=Llama-2-7b-Chat, Harmful span length=252026.06 | 3.8 | |
| InsertionBackbone=Llama-2-7b-Chat, Harmful span length=252026.06 | 6.5 | |
| PrefillBackbone=Llama-2-7b-Chat, Harmful span length=502026.06 | 9 | |
| InsertionBackbone=Llama-2-7b-Chat, Harmful span length=502026.06 | 11 | |
| PrefillBackbone=Llama-2-7b-Chat, Harmful span length=752026.06 | 13.2 | |
| InsertionBackbone=Llama-2-7b-Chat, Harmful span length=752026.06 | 13.5 | |
| PrefillBackbone=Llama-2-7b-Chat, Harmful span length=1002026.06 | 20.9 | |
| InsertionBackbone=Llama-2-7b-Chat, Harmful span length=1002026.06 | 21.2 | |
| CleanBackbone=Llama-2-7b-Chat, Harmful span length=252026.06 | 75.3 | |
| CleanBackbone=Llama-2-7b-Chat, Harmful span length=502026.06 | 80.4 | |
| CleanBackbone=Llama-2-7b-Chat, Harmful span length=752026.06 | 81.6 | |
| CleanBackbone=Llama-2-7b-Chat, Harmful span length=1002026.06 | 84.3 |