README.md

Agent Gym

Convert any model into a r1-like reasoning hyper-intelligent agent. Leverages TRL, Huggingface, and various other libraries. This is a work in progress. Our goal is to make it easy to train any model into a reasoning agent.

• Sources:
• Open R1 Blog
• GRPO Documentation from trl
• Huggingface Docs
• GRPO Docs

Installation

pip3 install -U agentgym

Usage

from agentgym.r1_pipeline import R1Pipeline, SFTConfig

r1_pipeline = R1Pipeline(
    sft_model="Qwen/Qwen2-0.5B-Instruct",
    tokenizer_name="Qwen/Qwen2-0.5B-Instruct",
    sft_dataset="trl-lib/tldr",
    sft_args=SFTConfig(output_dir="/tmp"),
    only_grpo=True,
    model_name="Qwen/Qwen2-0.5B-Instruct"
)

r1_pipeline.run()

Architecture

The architecture is as follows:

• SFT: Supervised Fine-Tuning
• GRPO: Generative Reinforcement Policy Optimization

-> model -> sft -> grpo -> model

graph TD;
    A[model] --> B[sft]
    B --> C[grpo]
    C --> D[reasoning model]

Loading

License

MIT