General Utility Evaluation on MT-Bench (MTB Score)
7.33MTB ScoreDPO mix50/1000/aug
Evaluation Results
| Method | Links | |
|---|---|---|
| DPO mix50/1000/augModel=Gemma-3-4B2026.05 | 7.33 | |
| DPO mix50/1000/augModel=Qwen3-4B2026.05 | 7.19 | |
| baselineModel=Gemma-3-4B2026.05 | 7.1 | |
| baselineModel=Qwen3-4B2026.05 | 7.04 | |
| baselineModel=Qwen2.5-7B2026.05 | 7.03 | |
| SFT mix0/200/cleanModel=Qwen2.5-7B2026.05 | 7.03 | |
| DPO mix50/1000/cleanModel=Qwen2.5-7B2026.05 | 7 | |
| SFT mix50/1000/augModel=Qwen3-4B2026.05 | 6.9 | |
| baselineModel=Llama-3.1-8B2026.05 | 6.82 | |
| DPO mix50/1000/augModel=Llama-3.1-8B2026.05 | 6.72 | |
| SFT mix50/200/augModel=Llama-3.1-8B2026.05 | 6.59 | |
| SFT mix50/1000/augModel=Gemma-3-4B2026.05 | 6.47 | |
| baselineModel=Mistral-7B2026.05 | 5.94 | |
| DPO mix50/1000/augModel=Mistral-7B2026.05 | 5.88 | |
| SFT mix50/1000/augModel=Mistral-7B2026.05 | 5.19 |