datawhalechina/agentic-rl-lab

Reproducing and studying RL algorithms for LLM agents, including GRPO, GSPO, DAPO, OPD, Search-R1, ReTool, ALFWorld and beyond.

What it solves

This project addresses the high barrier to entry for studying and implementing Agentic Reinforcement Learning (RL) algorithms. It removes common obstacles such as complex, tightly-coupled codebases and the need for massive GPU resources by providing simplified, reproducible code for cutting-edge research.

How it works

The repository acts as an experimental lab that breaks down complex RL algorithms into three parts: clear conceptual explanations, runnable code implementations (including data, rewards, loss functions, and training loops), and experiment tracking via SwanLab. It leverages the PyTRIO platform to enable remote training, allowing users to run experiments with lower hardware requirements.

Who it’s for

It is designed for researchers, algorithm engineers, and students who want to study, reproduce, or build upon state-of-the-art Agentic RL methods. It is particularly useful for those looking to develop projects or papers in the LLM and Agent RL space.

Highlights

  • Reproductions of advanced algorithms including GRPO, Search-R1, ReTool, DAPO, and OPSD.
  • Comprehensive documentation that explains the origin, purpose, and core variables of each algorithm.
  • Ready-to-use training pipelines for specific tasks like math problem-solving (GSM8K), medical capability enhancement, and household agent training (ALFWorld).
  • Integrated experiment tracking and modular code that can be easily adapted into personal projects.

Related

  • Project
    wilsonrljr/sysidentpy

    A Python library for system identification and time series forecasting using NARMAX models, offering tools for model structure selection and parameter estimation.

  • Project
    RWTHVision/diffusion-e2e-ft

    A framework for end-to-end fine-tuning of image-conditional diffusion models to create fast, single-step deterministic estimators for depth and surface normals.

  • Project
    koaning/embetter

    A library of scikit-learn compatible embeddings for text and vision that allows users to easily integrate pretrained encoders into ML pipelines.

  • Project
    aws/sagemaker-training-toolkit

    A library that makes custom Docker containers compatible with Amazon SageMaker, simplifying the process of training machine learning models in isolated environments.