lsdefine/simple_GRPO
A minimal GRPO implementation for training LLMs with reinforcement learning to achieve r1-style reasoning capabilities.
Not currently ranked — collecting fresh signals.
star history
This repository provides a simple implementation of Group Relative Policy Optimization (GRPO) for training large language models to exhibit reasoning behaviors similar to r1-style models. It includes support for vLLM inference acceleration, split reference models across GPUs, and memory-efficient training on single A800 GPUs. The codebase is designed for educational purposes and experimentation with RL training pipelines for LLMs.
Frequently asked
- What is lsdefine/simple_GRPO?
- A minimal GRPO implementation for training LLMs with reinforcement learning to achieve r1-style reasoning capabilities.
- Is simple_GRPO open source?
- Yes — lsdefine/simple_GRPO is open source, released under the Apache-2.0 license.
- What language is simple_GRPO written in?
- lsdefine/simple_GRPO is primarily written in Python.
- How popular is simple_GRPO?
- lsdefine/simple_GRPO has 1.7k stars on GitHub.
- Where can I find simple_GRPO?
- lsdefine/simple_GRPO is on GitHub at https://github.com/lsdefine/simple_GRPO.