Safety Evaluation on XSTest Safe (Refusal Rate)
51.6Refusal RateTarget feature
Evaluation Results
| Method | Links | |
|---|---|---|
| Target featureModel=Qwen3-4B-it2026.06 | 51.6 | |
| Target featureModel=Qwen3-1.7B-it2026.06 | 9.6 | |
| No SteeringModel=Qwen3-1.7B-it2026.06 | 4.4 | |
| No SteeringModel=Qwen3-4B-it2026.06 | 3.2 | |
| Random featureModel=Qwen3-4B-it2026.06 | 3.1 | |
| Random featureModel=Qwen3-1.7B-it2026.06 | 0 |