Toxicity Mitigation on Specialized category Manually-designed jailbreak attacks
4.5RTROptimus (CH)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Optimus (CH)Defense setting=Optimus (CH), Backbone=LLaMA-22025.07 | 4.5 | 5.85 | 9.9 | — | |
| No attackDefense setting=No attack, Backbone=LLaMA-22025.07 | 13.1 | 5.46 | 9.7 | — | |
| JB attackDefense setting=JB attack, Backbone=LLaMA-22025.07 | 58.3 | 6.01 | 9.7 | 3.21 |