zhangxy-2019/critique-GRPO ? reverse-engineered prompt

Build me a working version of this Critique GRPO project so I can train a language model that learns from both numeric rewards and natural language critiques.

I want the repo set up so I can create the training environment, start the training run, and reproduce the main workflow from the paper with the provided math training data and Qwen based setup. Please make sure the training script actually runs cleanly from the command line, and that the code is organized so it is easy to launch, monitor, and adjust if needed.

If anything important is missing or depends on a current library version, please check the latest docs online and update the code so it works in a fresh environment. I also want any obvious rough edges fixed, like broken paths, missing configs, or unclear setup steps, so I can go from clone to training without having to dig through the code.