Safety and Utility Evaluation on BeaverTails & LMSYS-Chat
97.88Rule ScoreShadow-Level
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Shadow-LevelBackbone=Qwen2.5-7B, Malicious client ratio=30%2026.01 | 97.88 | 96.92 | -0.87 | 5.17 | |
| Step-LevelBackbone=Qwen2.5-7B, Malicious client ratio=30%2026.01 | 70.38 | 58.46 | -2.42 | 5.04 | |
| Client-LevelBackbone=Qwen2.5-7B, Malicious client ratio=30%2026.01 | 69.04 | 56.73 | -2.52 | 5.14 |