Harmful Response Generation on HEx-PHI and AdvBench (test)
96.86ASRHarmRLVR
Evaluation results
| Method | Links | |
|---|---|---|
| HarmRLVRModel=Llama-3-8B-Instruct2025.10 | 96.86 | |
| SFTModel=Llama-3-8B-Instruct2025.10 | 84.28 | |
| ReNeLLMModel=Llama-3-8B-Instruct2025.10 | 70 | |
| CodeAttackModel=Llama-3-8B-Instruct2025.10 | 46 | |
| DPOModel=Llama-3-8B-Instruct2025.10 | 42.41 | |
| GCGModel=Llama-3-8B-Instruct2025.10 | 34.5 | |
| PAIRModel=Llama-3-8B-Instruct2025.10 | 18.7 |