Over-refusal Evaluation on WildGuard Unharmful
1.06Over-refusal RateCategorical Steering
Evaluation Results
| Method | Links | |
|---|---|---|
| Categorical SteeringBackbone=REFUSE-LLAMA, Method=Categorical Steering2026.03 | 1.06 | |
| Low-Rank CombinationBackbone=REFUSE-LLAMA, Method=Low-Rank Combination2026.03 | 3.81 | |
| Low-Rank CombinationBackbone=LLAMA 3 8B INSTRUCT, Method=Low-Rank Combination2026.03 | 5.08 | |
| REFUSE-LLAMABackbone=REFUSE-LLAMA2026.03 | 9.52 | |
| LLAMA 3 8B INSTRUCTBackbone=LLAMA 3 8B INSTRUCT2026.03 | 11.15 | |
| Low-Rank CombinationBackbone=DEEPSEEK R1 DISTILL LLAMA, Method=Low-Rank Combination2026.03 | 33.65 | |
| DEEPSEEK R1 DISTILL LLAMABackbone=DEEPSEEK R1 DISTILL LLAMA2026.03 | 39.05 |