Safety Evaluation on WildGuard (Unsafe)
54.9Refusal RateTarget feature
Evaluation Results
| Method | Links | |
|---|---|---|
| Target featureModel=Qwen3-4B-it2026.06 | 54.9 | |
| No SteeringModel=Qwen3-4B-it2026.06 | 40.5 | |
| Target featureModel=Qwen3-1.7B-it2026.06 | 40.1 | |
| No SteeringModel=Qwen3-1.7B-it2026.06 | 26.3 | |
| Random featureModel=Qwen3-1.7B-it2026.06 | 0.9 | |
| Random featureModel=Qwen3-4B-it2026.06 | 0 |