Mathematical Reasoning on AMC 2023 (test) (Accuracy avg@32)
93.8Accuracy (avg@32)NFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| NFTBackbone=Qwen2.5-32B, Fine-tuning Protocol=Supervised fine-tuning2025.05 | 93.8 | — | |
| RFTBackbone=Qwen2.5-32B, Fine-tuning Protocol=Supervised fine-tuning2025.05 | 92.4 | — | |
| DAPOBackbone=Qwen2.5-32B, Fine-tuning Protocol=Reinforcement fine-tuning2025.05 | 90.9 | — | |
| NFTBackbone=Qwen2.5-Math-7B, Fine-tuning Protocol=Supervised fine-tuning2025.05 | 88.5 | — | |
| DAPOBackbone=Qwen2.5-Math-7B, Fine-tuning Protocol=Reinforcement fine-tuning2025.05 | 85 | — | |
| DPOBackbone=Qwen2.5-Math-7B, Fine-tuning Protocol=Preference fine-tuning2025.05 | 83.2 | — | |
| Dr. GRPOBackbone=Qwen2.5-Math-7B, Fine-tuning Protocol=Reinforcement fine-tuning2025.05 | 80.2 | — | |
| RFTBackbone=Qwen2.5-Math-7B, Fine-tuning Protocol=Supervised fine-tuning2025.05 | 79.7 | — | |
| GRPOBackbone=Qwen2.5-Math-7B, Fine-tuning Protocol=Reinforcement fine-tuning2025.05 | 79.5 | — | |
| Qwen2.5-Math-7BBackbone=Qwen2.5-Math-7B, Fine-tuning Protocol=Base2025.05 | 45.8 | — | |
| Qwen2.5-32BBackbone=Qwen2.5-32B, Fine-tuning Protocol=Base2025.05 | 45 | — | |
| GRPOBackbone=Qwen3-1.7B, Training Step Selection=Last (Step 400)2026.06 | — | 46.7 | |
| GRPOBackbone=Qwen3-1.7B, Training Step Selection=Best (Peak)2026.06 | — | 46.7 | |
| OPCDBackbone=Qwen3-1.7B, Training Step Selection=Last (Step 400)2026.06 | — | 24.5 | |
| OPCDBackbone=Qwen3-1.7B, Training Step Selection=Best (Peak)2026.06 | — | 47.4 | |
| RLSDBackbone=Qwen3-1.7B, Training Step Selection=Last (Step 400)2026.06 | — | 57.3 | |
| RLSDBackbone=Qwen3-1.7B, Training Step Selection=Best (Peak)2026.06 | — | 58.2 | |
| SDPG-UFKLBackbone=Qwen3-1.7B, Training Step Selection=Last (Step 400)2026.06 | — | 63.7 | |
| SDPG-UFKLBackbone=Qwen3-1.7B, Training Step Selection=Best (Peak)2026.06 | — | 66.6 | |
| SDPG-URKLBackbone=Qwen3-1.7B, Training Step Selection=Last (Step 400)2026.06 | — | 62 | |
| SDPG-URKLBackbone=Qwen3-1.7B, Training Step Selection=Best (Peak)2026.06 | — | 62 |