中文版

• multiturn_grpo.yaml and multiturn_megatron_grpo.yaml are configuration files under the Verl framework, corresponding to the standard version and Megatron version of the GRPO model respectively.
• multiturn_llm_reward.py is a custom reward model file under the Verl framework used for calculating reward values.

Usage instructions:

1. Ensure that the Verl framework and related dependencies are installed.
2. Point the model path in the multiturn_grpo.yaml or multiturn_megatron_grpo.yaml configuration file to the pretrained GRPO model.
3. Point the reward model path in the multiturn_llm_reward.py configuration file to this file.
4. Run the configuration file using the Verl framework to start training and evaluating the model.