Harmfulness Evaluation on I-Malicious
0Harmful RateNLCf/800 step
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| NLCf/800 stepSteps=800, Loss=NLCf2026.05 | 0 | 0.2 | |
| EMD/800 stepSteps=800, Loss=EMD2026.05 | 0 | 0.16 | |
| Mistral MoE-XLArchitecture=MoE-XL2026.05 | 6 | 0.005 | |
| Qwen-3B MoE-XLArchitecture=MoE-XL2026.05 | 13 | 0.01 | |
| Mistral2026.05 | 27 | 3 | |
| Qwen-3B2026.05 | 34 | 0.03 |