Multi-turn Conversation on MT-Bench (Win Rate)
77.1Win RateRPO
Evaluation Results
| Method | Links | |
|---|---|---|
| RPOModel=Qwen2.5-7B, Baseline=SFT, LLM Judge=GPT-4o2025.06 | 77.1 | |
| RPOModel=Qwen2.5-7B, Baseline=SFT, LLM Judge=Claude-3.52025.06 | 76.8 | |
| RPOModel=LLaMA-3-8B, Baseline=SFT, LLM Judge=GPT-4o2025.06 | 76.4 | |
| RPOModel=LLaMA-3-8B, Baseline=SFT, LLM Judge=Claude-3.52025.06 | 75.9 | |
| RPOModel=Mistral-7B, Baseline=SFT, LLM Judge=GPT-4o2025.06 | 75.1 | |
| RPOModel=Mistral-7B, Baseline=SFT, LLM Judge=Claude-3.52025.06 | 74.6 | |
| RPOModel=FlanT5-XL, Baseline=SFT, LLM Judge=GPT-4o2025.06 | 68.2 | |
| RPOModel=FlanT5-XL, Baseline=SFT, LLM Judge=Claude-3.52025.06 | 67.4 | |
| RPOModel=FlanT5-Large, Baseline=SFT, LLM Judge=GPT-4o2025.06 | 64.3 | |
| RPOModel=FlanT5-Large, Baseline=SFT, LLM Judge=Claude-3.52025.06 | 63.8 | |
| RPOModel=Qwen2.5-7B, Baseline=KTO, LLM Judge=GPT-4o2025.06 | 61.5 | |
| RPOModel=LLaMA-3-8B, Baseline=KTO, LLM Judge=GPT-4o2025.06 | 61 | |
| RPOModel=Qwen2.5-7B, Baseline=KTO, LLM Judge=Claude-3.52025.06 | 60.9 | |
| RPOModel=LLaMA-3-8B, Baseline=KTO, LLM Judge=Claude-3.52025.06 | 60.3 | |
| RPOModel=Mistral-7B, Baseline=KTO, LLM Judge=GPT-4o2025.06 | 60.2 | |
| RPOModel=Mistral-7B, Baseline=KTO, LLM Judge=Claude-3.52025.06 | 59.1 | |
| RPOModel=FlanT5-Small, Baseline=SFT, LLM Judge=GPT-4o2025.06 | 58.8 | |
| RPOModel=FlanT5-Small, Baseline=SFT, LLM Judge=Claude-3.52025.06 | 57.9 | |
| RPOModel=FlanT5-XL, Baseline=KTO, LLM Judge=GPT-4o2025.06 | 57.4 | |
| RPOModel=FlanT5-XL, Baseline=KTO, LLM Judge=Claude-3.52025.06 | 56.9 | |
| RPOModel=Qwen2.5-7B, Baseline=DRO, LLM Judge=GPT-4o2025.06 | 56.8 | |
| RPOModel=LLaMA-3-8B, Baseline=DRO, LLM Judge=GPT-4o2025.06 | 56.2 | |
| RPOModel=Qwen2.5-7B, Baseline=DRO, LLM Judge=Claude-3.52025.06 | 56.1 | |
| RPOModel=FlanT5-Large, Baseline=KTO, LLM Judge=GPT-4o2025.06 | 56 | |
| RPOModel=Mistral-7B, Baseline=DRO, LLM Judge=GPT-4o2025.06 | 55.9 | |
| RPOModel=LLaMA-3-8B, Baseline=DRO, LLM Judge=Claude-3.52025.06 | 55.8 | |
| RPOModel=FlanT5-Large, Baseline=KTO, LLM Judge=Claude-3.52025.06 | 55.4 | |
| RPOModel=Mistral-7B, Baseline=DRO, LLM Judge=Claude-3.52025.06 | 55.2 | |
| RPOModel=FlanT5-Small, Baseline=KTO, LLM Judge=GPT-4o2025.06 | 53.7 | |
| RPOModel=FlanT5-XL, Baseline=DRO, LLM Judge=GPT-4o2025.06 | 53.5 | |
| RPOModel=FlanT5-Small, Baseline=KTO, LLM Judge=Claude-3.52025.06 | 52.9 | |
| RPOModel=FlanT5-Large, Baseline=DRO, LLM Judge=GPT-4o2025.06 | 52.9 | |
| RPOModel=FlanT5-XL, Baseline=DRO, LLM Judge=Claude-3.52025.06 | 52.8 | |
| RPOModel=FlanT5-Large, Baseline=DRO, LLM Judge=Claude-3.52025.06 | 52.1 | |
| RPOModel=FlanT5-Small, Baseline=DRO, LLM Judge=GPT-4o2025.06 | 51.5 | |
| RPOModel=FlanT5-Small, Baseline=DRO, LLM Judge=Claude-3.52025.06 | 50.8 |