#1111th of 41 in Training and fine-tuning
art
Python RL framework that trains LLM agents with GRPO from rollout rewards
Documentation ↗
(opens in a new tab)Website ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 10.8k
- License
- Apache-2.0
- Last commit
- Oct 2026
- Last release
- Oct 2026
- Language
- Python
9,059 on 10 Apr 202610,836 on 30 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
ART wraps GRPO reinforcement learning around existing Python agent code. An OpenAI-compatible client runs rollouts and records trajectories with rewards, while a separate server runs vLLM inference and trains LoRAs on a GPU. It can also use W&B's hosted Serverless RL backend instead of local GPUs.
Who it is for: Engineers fine-tuning LLM agents with reinforcement learning
Strengths
- Client/server split: run the client on a laptop, train on a remote or local GPU
- OpenAI-compatible client drops into existing agent code, with a LangGraph integration
- RULER automatic reward generation and notebooks for 2048, email search, MCP and SFT
- Optional W&B Serverless RL backend avoids managing GPU infrastructure
Weaknesses
- Inference is blocked while each training step runs
- Gemma 3 does not appear to be supported
- You must define rewards and rollout environments yourself, unless using RULER
- Serverless path depends on a W&B account and API key
What it needs
- GPU required
- Docker
- Needs vLLM, Unsloth, W&B (optional, Serverless RL)
- Models: vLLM/HuggingFace causal LMs, Qwen, LoRA
Quick answers
Is art open source?
Yes. Its code is on GitHub under the Apache-2.0 license.
Can you run art with Docker?
Yes. Its repository has a Docker setup: a Dockerfile, a Compose file or a published image.
Does art need a GPU?
Yes, it needs a GPU.
Also in Training and fine-tuning
See all 41| Rank | Project | Score |
|---|---|---|
| 1 | unslothDesktop and web app to run and fine-tune local models | 88 out of 100 |
| 2 | peftPython library for LoRA and other parameter-efficient fine-tuning of Transformers models | 86 out of 100 |
| 3 | deepspeedPyTorch library for distributed training of very large models with ZeRO | 85 out of 100 |
| #4 | verlReinforcement learning training library for LLM post-training with PPO and GRPO | 84 out of 100 |
| #5 | trlPost-training library with SFT, DPO, GRPO and KTO trainers | 83 out of 100 |