Safety Evaluation on WildGuard
25.5WildGuard Refusal RateBase
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BaseModel=Mistral-7B2026.05 | 25.5 | 59.5 | |
| BaseModel=Qwen3-0.6B2026.05 | 32 | 52 | |
| TrigModel=Qwen3-0.6B, Sparsity=81.05%2026.05 | 39.5 | 41.5 | |
| BaseModel=DeepSeek-8B2026.05 | 47.5 | 38.5 | |
| TrigModel=DeepSeek-8B, Sparsity=81.38%2026.05 | 56.5 | 30.5 | |
| TrigModel=Mistral-7B, Sparsity=59.34%2026.05 | 69 | 21.5 | |
| TrigModel=Vicuna-7B, Sparsity=84.40%2026.05 | 74.5 | 23 | |
| BaseModel=Vicuna-7B2026.05 | 76.5 | 20.5 | |
| LCDDModel=Qwen3-0.6B, Sparsity=81.05%2026.05 | 88.5 | 3.5 | |
| SFTModel=Qwen3-0.6B2026.05 | 94 | 2 | |
| LCDDModel=DeepSeek-8B, Sparsity=81.38%2026.05 | 96.5 | 2 | |
| LCDDModel=Mistral-7B, Sparsity=59.34%2026.05 | 97.5 | 1 | |
| LCDDModel=Vicuna-7B, Sparsity=84.40%2026.05 | 98 | 1.5 | |
| SFTModel=Vicuna-7B2026.05 | 99 | 0 | |
| SFTModel=DeepSeek-8B2026.05 | 100 | 0 | |
| SFTModel=Mistral-7B2026.05 | 100 | 0 |