Safety and Utility Evaluation on MaliciousGen & WildChat
97.69Rule AdherenceStep-Level
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Step-LevelBackbone=Qwen2.5-7B, Malicious client ratio=30%2026.01 | 97.69 | 95.38 | -0.8 | 5.24 | |
| Shadow-LevelBackbone=Qwen2.5-7B, Malicious client ratio=30%2026.01 | 97.69 | 96.35 | -0.77 | 5.68 | |
| Client-LevelBackbone=Qwen2.5-7B, Malicious client ratio=30%2026.01 | 97.31 | 95.77 | -0.76 | 5.33 |