Instruction Following Evaluation on Alpaca-Eval
62.17Length-Controlled Win RateFiMi-RM
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| FiMi-RMBase Model=Qwen2.5-7B, Evaluation Algorithm=DPO2025.05 | 62.17 | 67.32 | 757 | 174 | |
| FiMi-RMBase Model=Gemma2-9B, Evaluation Algorithm=DPO2025.05 | 59.47 | 61.61 | 773 | 189 | |
| ODINBase Model=Qwen2.5-7B, Evaluation Algorithm=DPO2025.05 | 58.91 | 63.16 | 1,044 | 244 | |
| Vanilla RMBase Model=Qwen2.5-7B, Evaluation Algorithm=DPO2025.05 | 58.56 | 61.73 | 1,089 | 254 | |
| ODINBase Model=Gemma2-9B, Evaluation Algorithm=DPO2025.05 | 57.96 | 61.19 | 905 | 214 | |
| Length PenaltyBase Model=Gemma2-9B, Evaluation Algorithm=DPO2025.05 | 57.78 | 60.03 | 1,077 | 250 | |
| Length PenaltyBase Model=Qwen2.5-7B, Evaluation Algorithm=DPO2025.05 | 57.67 | 62.89 | 980 | 227 | |
| Vanilla RMBase Model=Gemma2-9B, Evaluation Algorithm=DPO2025.05 | 57.35 | 59.81 | 1,135 | 267 |