Reasoning on GSM8K Track I
88.02Overall GSM8K AccuracyGRPO+CLIPO
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GRPO+CLIPOGlobal batch size=512, Training Epochs=5, Max response length=20482026.03 | 88.02 | 84.62 | 74.6 | 54.16 | 76.9 | 31.52 | 27.58 | 68.64 | 75.35 | 51.16 | 63.26 | |
| GSPO+CLIPOGlobal batch size=512, Training Epochs=5, Max response length=20482026.03 | 87.95 | 83.98 | 74.54 | 52.92 | 75.92 | 32.19 | 26.1 | 67.48 | 74.85 | 50.42 | 62.63 | |
| GRPOGlobal batch size=512, Training Epochs=5, Max response length=20482026.03 | 87.79 | 84.04 | 73.12 | 50.8 | 74.94 | 30.48 | 27.71 | 68.19 | 73.94 | 50.33 | 62.13 | |
| GMPO+CLIPOGlobal batch size=512, Training Epochs=5, Max response length=20482026.03 | 87.64 | 85.52 | 75.52 | 50.92 | 75.35 | 31.95 | 27.58 | 67.63 | 74.9 | 50.63 | 62.76 | |
| DAPO+CLIPOGlobal batch size=512, Training Epochs=5, Max response length=20482026.03 | 87.57 | 84.66 | 74.54 | 53.12 | 77.07 | 31.4 | 27.44 | 67.93 | 74.97 | 50.96 | 62.97 | |
| DAPOGlobal batch size=512, Training Epochs=5, Max response length=20482026.03 | 87.41 | 84.56 | 74.7 | 52.32 | 76 | 31.58 | 27.11 | 69 | 74.75 | 50.92 | 62.84 | |
| GSPOGlobal batch size=512, Training Epochs=5, Max response length=20482026.03 | 87.19 | 83.94 | 73.28 | 51.36 | 74.86 | 30.23 | 27.38 | 67.68 | 73.94 | 50.04 | 61.99 | |
| GMPOGlobal batch size=512, Training Epochs=5, Max response length=20482026.03 | 86.81 | 84.02 | 73.9 | 49.8 | 76.33 | 30.97 | 26.57 | 68.49 | 73.63 | 50.59 | 62.11 | |
| Base ModelGlobal batch size=512, Training Epochs=5, Max response length=20482026.03 | 86.05 | 82.4 | 70.9 | 47.96 | 73.63 | 29.99 | 24.9 | 68.09 | 71.83 | 49.15 | 60.49 |