Question for selection of experiments results

[Trae1ounG]

Great work!

Could you please clarify whether the performance results for MATH-500, AMC, and AIME 2024 (with the number of episodes set to 10, 30, and 80, respectively) are based on selecting the last-step checkpoints or the best results during training?

While reproducing GRPO on AIME 2024, I observed that the last-step avg@16 performance is 13.13, whereas the best performance during training can exceed 17.

Thank you for your assistance!

[Gank0078]

The same problem. I cannot reproduce the results in the paper, such as Qwen2.5 and Qwe2.5 Math 1.5B.