Bias Evaluation on Race & SES
17Mean ImprovementiPASwo
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| iPASwoModel=Llama-3.1-8B-Instruct, Steering Target=residual2025.09 | 17 | 0.16 | 0 | |
| iPASwoModel=Nous-Hermes-2-Mistral-7B-DPO, Steering Target=residual2025.09 | 10 | 0.08 | 0 | |
| iPASwoModel=Llama-3.1-8B-Instruct, Steering Target=residual2025.09 | 10 | 0.09 | 0 | |
| iPASwoModel=DeepSeek-R1-Distill-8B, Steering Target=residual2025.09 | 10 | 0.09 | 0 | |
| iPASaModel=Llama-3.1-8B-Instruct, Steering Target=residual2025.09 | 10 | 0.09 | 0 | |
| iPASwoModel=DeepSeek-R1-Distill-8B, Steering Target=residual2025.09 | 9 | 0.08 | 0 | |
| iPASaModel=Llama-3.1-8B-Instruct, Steering Target=residual2025.09 | 9 | 0.08 | 0 | |
| iPASaModel=DeepSeek-R1-Distill-8B, Steering Target=residual2025.09 | 8 | 0.07 | 0 | |
| iPASwoModel=Nous-Hermes-2-Mistral-7B-DPO, Steering Target=residual2025.09 | 8 | 0.07 | 0 | |
| iPASaModel=Nous-Hermes-2-Mistral-7B-DPO, Steering Target=residual2025.09 | 7 | 0.06 | 0 | |
| iPASaModel=Nous-Hermes-2-Mistral-7B-DPO, Steering Target=residual2025.09 | 7 | 0.06 | 0 | |
| iPASaModel=DeepSeek-R1-Distill-8B, Steering Target=residual2025.09 | 7 | 0.07 | 0 | |
| PASfModel=Nous-Hermes-2-Mistral-7B-DPO, Steering Target=residual2025.09 | 5 | 0.04 | 0 | |
| PASfModel=Llama-3.1-8B-Instruct, Steering Target=residual2025.09 | 4 | 0.03 | 0 | |
| PASfModel=Llama-3.1-8B-Instruct, Steering Target=residual2025.09 | 4 | 0.03 | 0 | |
| PASfModel=Nous-Hermes-2-Mistral-7B-DPO, Steering Target=residual2025.09 | 3 | 0.02 | 0 | |
| PASfModel=DeepSeek-R1-Distill-8B, Steering Target=residual2025.09 | 1 | 0 | 0 | |
| PASfModel=DeepSeek-R1-Distill-8B, Steering Target=residual2025.09 | 0 | -0.01 | 0.61 |