This is the step 2720 checkpoint when training Qwen2.5-Math-1.5B on MATH Level-3-to-5 Dataset PrAg-PO. The training procedure is outlined in the paper
PrAg-PO: Prompt Augmented Policy Optimization for Robust and Diverse Mathematical Reasoning.
This model is intended for mathematical reasoning tasks. It leverages prompt augmentation to generate reasoning traces under diverse templates, increasing rollout diversity and stability during RL training.
1@misc{lu2026pragpopromptaugmentedpolicy,
2 title={PrAg-PO: Prompt Augmented Policy Optimization for Robust and Diverse Mathematical Reasoning},
3 author={Wenquan Lu and Hai Huang and Enqi Liu and Randall Balestriero},
4 journal={arXiv preprint arXiv:2602.03190},
5 url={https://arxiv.org/abs/2602.03190},
6 year={2026},
7}
