Structural Bias Evaluation on MNLI
98.1AccuracyC2PO
Evaluation Results
| Method | Links | |
|---|---|---|
| C2POBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 98.1 | |
| GRPOBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 92.5 | |
| C2POBackbone=LLaMA-2-13B-Chat2025.12 | 86.2 | |
| CPOBackbone=LLaMA-2-13B-Chat2025.12 | 84.1 | |
| IPOBackbone=LLaMA-2-13B-Chat2025.12 | 81.2 | |
| FRBackbone=LLaMA-2-13B-Chat2025.12 | 80.3 | |
| FRBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 79.2 | |
| DPOBackbone=LLaMA-2-13B-Chat2025.12 | 67.2 | |
| CPOBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 65.1 | |
| IPOBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 64.8 | |
| BCOBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 64.7 | |
| BCOBackbone=LLaMA-2-13B-Chat2025.12 | 64.4 | |
| GRPOBackbone=LLaMA-2-13B-Chat2025.12 | 58.4 | |
| DPOBackbone=DeepSeek-R1-Distill-Qwen-14B2025.12 | 37.9 |