Safety Evaluation on BeaverTails Self Harm
100Safe RateSafeMoE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SafeMoEBase Model=Mistral-7B, Variant=MoE-XL2026.05 | 100 | 6.26 | |
| SafeMoEBase Model=Qwen-3B, Variant=MoE-XL2026.05 | 87.5 | 7.86 | |
| Qwen-3BTuning=Standard instruction-tuned2026.05 | 33.3 | 7.5 | |
| Mistral-7BTuning=Standard instruction-tuned2026.05 | 18.8 | 5 |