Classification on MNLI
33.1ASRPROTOPURIFY
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PROTOPURIFYAttack=VPI, Backbone=Mistral-7B2026.02 | 33.1 | 46.2 | |
| PROTOPURIFYAttack=CBA, Backbone=Mistral-7B2026.02 | 14.9 | 86.8 | |
| PROTOPURIFYAttack=CBA, Backbone=Llama3-8B2026.02 | 14 | 86.9 | |
| PROTOPURIFYAttack=BadEdit, Backbone=Llama3-8B2026.02 | 10.1 | 64.3 | |
| PROTOPURIFYAttack=VPI, Backbone=Llama3-8B2026.02 | 8.5 | 45.8 | |
| PROTOPURIFYAttack=BadNet, Backbone=Mistral-7B2026.02 | 5.6 | 81.7 | |
| PROTOPURIFYAttack=BadEdit, Backbone=Mistral-7B2026.02 | 4.1 | 49.7 | |
| PROTOPURIFYAttack=BadNet, Backbone=Llama3-8B2026.02 | 1.6 | 83.1 |