Bias Evaluation (Age) on BBQ Age (test)
92.5AccuracyKP
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| KPModel=Qwen 2.5 7B2025.09 | 92.5 | — | 100 | 0 | |
| KPModel=Llama 3.1 8B2025.09 | 89.1 | — | 100 | 0 | |
| KAPPAModel=Qwen 2.5 7B, Intervention layers=62025.09 | 85.5 | — | 87 | 0.64 | |
| CAAModel=Qwen 2.5 7B2025.09 | 84.3 | — | 85 | 0.67 | |
| KAPPAModel=Qwen 2.5 7B, Intervention layers=12025.09 | 84.1 | — | 85.2 | 0.67 | |
| KAPPAModel=Qwen 2.5 7B, Intervention layers=32025.09 | 83.9 | — | 84.9 | 0.67 | |
| Qwen 2.5 7B Instructk=32025.09 | 83.2 | 9.3 | 84 | 0.68 | |
| BaseModel=Qwen 2.5 7B2025.09 | 83.2 | — | 84 | 0.68 | |
| DoLAModel=Qwen 2.5 7B2025.09 | 82.9 | — | 83.9 | 0.68 | |
| KAPPAModel=Llama 3.1 8B, Intervention layers=62025.09 | 76.8 | — | 81.1 | 0.31 | |
| KAPPAModel=Llama 3.1 8B, Intervention layers=12025.09 | 73.8 | — | 75 | 0.45 | |
| KAPPAModel=Llama 3.1 8B, Intervention layers=32025.09 | 72.3 | — | 74.4 | 0.38 | |
| CAAModel=Llama 3.1 8B2025.09 | 65.8 | — | 64.8 | 0.53 | |
| Llama 3.1 8B Instructk=32025.09 | 59.9 | 29.3 | 59.2 | 0.59 | |
| BaseModel=Llama 3.1 8B2025.09 | 59.9 | — | 59.2 | 0.59 | |
| DoLAModel=Llama 3.1 8B2025.09 | 59.6 | — | 60.1 | 0.64 |