Stereotypical Bias Mitigation on Unqover
99.9AccuracyGRPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GRPOBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 99.9 | 0.2 | |
| C2POBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 99.9 | 0.2 | |
| CPOBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 99.7 | 1.9 | |
| IPOBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 99.6 | 0.7 | |
| BCOBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 96.3 | 5.1 | |
| FRBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 94.9 | 7 | |
| DPOBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 93.3 | 11.6 | |
| C2POBackbone=LLaMA-2-13B-Chat2025.12 | 91.8 | 3 | |
| CPOBackbone=LLaMA-2-13B-Chat2025.12 | 85.9 | 8.8 | |
| BCOBackbone=LLaMA-2-13B-Chat2025.12 | 85.3 | 9.4 | |
| IPOBackbone=LLaMA-2-13B-Chat2025.12 | 83.3 | 10.9 | |
| GRPOBackbone=LLaMA-2-13B-Chat2025.12 | 36.3 | 17.4 | |
| FRBackbone=LLaMA-2-13B-Chat2025.12 | 26.4 | 7.1 | |
| DPOBackbone=LLaMA-2-13B-Chat2025.12 | 23.3 | 7.4 |