Harmful Refusal on StrongREJECT vani harm
99.5RTA Rate (vani harm)Qwen2.5-14B-IT + SELF-REDTEAM
Evaluation Results
| Method | Links | |
|---|---|---|
| Qwen2.5-14B-IT + SELF-REDTEAMTraining=SELF-REDTEAM fine-tuned2025.06 | 99.5 | |
| Qwen2.5-7B-IT + SELF-REDTEAMTraining=SELF-REDTEAM fine-tuned2025.06 | 97.9 | |
| Qwen2.5-3B-IT + SELF-REDTEAMTraining=SELF-REDTEAM fine-tuned2025.06 | 97.1 | |
| Qwen2.5-14B-ITTraining=Instruction-tuned base2025.06 | 97.1 | |
| Llama3.1-8B-ITTraining=Instruction-tuned base2025.06 | 97.1 | |
| Llama3.1-8B-IT + SELF-REDTEAMTraining=SELF-REDTEAM fine-tuned2025.06 | 95.8 | |
| Qwen2.5-3B-ITTraining=Instruction-tuned base2025.06 | 92 | |
| Qwen2.5-7B-ITTraining=Instruction-tuned base2025.06 | 92 | |
| Llama3.1-8B-IT-AB + SELF-REDTEAMTraining=SELF-REDTEAM fine-tuned2025.06 | 91.2 | |
| Llama3.1-8B-IT-ABTraining=Instruction-tuned ablated base2025.06 | 12.1 |