Adversarial Toxicity Refusal on LLaMA-2 Chatbot Offensive category
47.7RTRAdv. attack
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Adv. attackDefense setting=Adv. attack2025.07 | 47.7 | 5.29 | 9.8 | -2.86 | |
| No attackDefense setting=No attack2025.07 | 8.8 | 4.27 | 10.2 | — | |
| Optimus (NH)Defense setting=Optimus (NH)2025.07 | 0.2 | 6.17 | 10.9 | — |