Conversational Skill Evaluation on Arena-Hard
32.6Win Rate (%)CPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CPOBackbone=Llama-3-8B-Instruct2026.05 | 32.6 | 1.9 | |
| SimPOBackbone=Llama-3-8B-Instruct2026.05 | 30 | 2.7 | |
| DPOBackbone=Llama-3-8B-Instruct2026.05 | 28.9 | 1.7 | |
| RDPOBackbone=Llama-3-8B-Instruct2026.05 | 28.5 | 1.9 | |
| IPOBackbone=Llama-3-8B-Instruct2026.05 | 26.9 | 1.6 | |
| Contrastive-POBackbone=Llama-3-8B-Instruct2026.05 | 26.6 | 1.6 | |
| KTOBackbone=Llama-3-8B-Instruct2026.05 | 23.1 | 1 | |
| SLiC-HFBackbone=Llama-3-8B-Instruct2026.05 | 22.9 | 1.4 | |
| ORPOBackbone=Llama-3-8B-Instruct2026.05 | 21.7 | 1.1 | |
| RRHFBackbone=Llama-3-8B-Instruct2026.05 | 20.5 | 1.5 | |
| SFT-BaseBackbone=Llama-3-8B-Instruct2026.05 | 19.2 | 1.4 |