Safety Evaluation on BeaverTails (Average)
87Safety ScoreSafeMoE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SafeMoEBase Model=Mistral-7B, Variant=MoE-XL2026.05 | 87 | 6.34 | |
| SafeMoEBase Model=Qwen-3B, Variant=MoE-XL2026.05 | 63.39 | 6.9 | |
| Qwen-3BTuning=Standard instruction-tuned2026.05 | 34.03 | 6.3 | |
| Mistral-7BTuning=Standard instruction-tuned2026.05 | 31.44 | 3.91 |