List your company

Safety and Utility Evaluation on BeaverTails & WildChat

97.5Rule AdherenceShadow-Level
43.898457.814271.7385.6458Jan 12, 2026

Evaluation results

MethodLinks
Shadow-LevelBackbone=Qwen2.5-7B, Malicious client ratio=30%2026.0197.595.96-0.85.39
Client-LevelBackbone=Qwen2.5-7B, Malicious client ratio=30%2026.0183.8575.77-1.155.49
Step-LevelBackbone=Qwen2.5-7B, Malicious client ratio=30%2026.0183.6572.12-1.345.13
Client-LevelBackbone=Llama3.1-8B, Malicious client ratio=30%, Base Framework=Safe-FedLLM2026.0182.1254.81-1.833.95
Step-LevelBackbone=Llama3.1-8B, Malicious client ratio=30%, Base Framework=Safe-FedLLM2026.0180.5852.5-1.893.02
Shadow-LevelBackbone=Llama3.1-8B, Malicious client ratio=30%, Base Framework=Safe-FedLLM2026.0179.4251.35-1.864.1
FoolsGoldBackbone=Llama3.1-8B, Malicious client ratio=30%2026.0151.547.69-3.183.91
ResidualBackbone=Llama3.1-8B, Malicious client ratio=30%2026.0150.587.31-3.324.14
FedAvgBackbone=Llama3.1-8B, Malicious client ratio=30%2026.0149.427.88-3.453.95
TrimmedMeanBackbone=Llama3.1-8B, Malicious client ratio=30%2026.0147.515.96-3.313.91
KrumBackbone=Llama3.1-8B, Malicious client ratio=30%2026.0145.964.42-3.374.3