Toxicity mitigation evaluation on Specialized category
59.6RTROptimus
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| OptimusBackbone=LLaMA-2, Defense setting=Attack2025.07 | 59.6 | 5.83 | 0.097 | 60.1 | |
| OptimusBackbone=LLaMA-2, Defense setting=No-attack2025.07 | 13.1 | 5.46 | 0.097 | 59.6 | |
| OptimusBackbone=LLaMA-2, Defense setting=O-API, Healing type=Contextual (CH)2025.07 | 9.7 | 6 | 0.099 | 61 | |
| OptimusBackbone=LLaMA-2, Defense setting=P-API, Healing type=Non-contextual (NH)2025.07 | 8.2 | 6.19 | 0.1 | 60.8 | |
| OptimusBackbone=LLaMA-2, Defense setting=P-API, Healing type=Contextual (CH)2025.07 | 7.7 | 6 | 0.1 | 61.5 | |
| OptimusBackbone=LLaMA-2, Defense setting=O-API, Healing type=Non-contextual (NH)2025.07 | 6.8 | 6.2 | 0.101 | 60.9 | |
| OptimusBackbone=LLaMA-2, Defense setting=Refusal, Healing type=Contextual (CH)2025.07 | 3.2 | 6.11 | 0.102 | 60.4 | |
| OptimusBackbone=LLaMA-2, Defense setting=Refusal, Healing type=Non-contextual (NH)2025.07 | 3 | 6.27 | 0.103 | 60.6 |