Adversarial Toxicity Refusal on LLaMA-2 Chatbot Specialized category
59.8Refusal Rate (RTR)Adv. attack
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Adv. attackDefense setting=Adv. attack2025.07 | 59.8 | 5.85 | 9.7 | — | |
| No attackDefense setting=No attack2025.07 | 13.1 | 5.46 | 9.7 | — | |
| Optimus (NH)Defense setting=Optimus (NH)2025.07 | 1.8 | 6.23 | 10.3 | -1.55 |