Over-refusal Mitigation on Gemma-2B Over-refusal
1.6Informative Refusal RateOriginal
Evaluation Results
| Method | Links | |
|---|---|---|
| OriginalInference Precision=FP322026.06 | 1.6 | |
| OriginalInference Precision=FP322026.06 | 2.4 | |
| OriginalInference Precision=FP322026.06 | 2.7 | |
| QVecInference Precision=LLM.int8()2026.06 | 3.4 | |
| GaussianInference Precision=FP42026.06 | 3.73 | |
| QVecInference Precision=NF42026.06 | 6.87 | |
| GaussianInference Precision=NF42026.06 | 7.07 | |
| GaussianInference Precision=LLM.int8()2026.06 | 7.2 | |
| QVecInference Precision=FP42026.06 | 10.3 | |
| AttackedInference Precision=LLM.int8()2026.06 | 25.9 | |
| AttackedInference Precision=NF42026.06 | 39.5 | |
| AttackedInference Precision=FP42026.06 | 60.7 |