math500

A small experiment repository comparing a base reasoning model against RLVR-GRPO checkpoints on the Math500 dataset. It includes evaluation results, short-form observations, and a local temp_clone of the full open-posttraining-system codebase for reference.

Auditable reproduction of Qwen2.5-7B vs SimpleRL-Zoo on MATH500 for Limit of RLVR (arXiv:2504.13837 v5).