🎨ART
Agent Reinforcement Trainer: train multi-step agents for real-world tasks using GRPO. Give your agents on-the-job training. Reinforcement learning for Qwen3.6, GPT-OSS, Llama, and more!
⭐ Very popular: 11k stars, gaining about 41 a week
View on GitHub ↗repo profile
momentum
durability
bus factor = how many people it takes to cover more than half the commits (6 months). 1 is a solo project; higher means the work is spread across a team. top-author share is the single busiest author's slice of those commits.
since we covered it
why it's a big deal
- Gives Python developers a way to improve LLM agents on multi-step, real-world tasks through reinforcement learning rather than prompt tweaking alone.
- RULER auto-generates rewards, so teams can run RL without hand-engineering a reward function for each task.
- Works with common open models including Qwen, Llama, and GPT-OSS, and integrates with LangGraph and MCP servers for agent workflows.
under the hood
- Trains with GRPO in a client-server loop where an OpenAI-compatible client runs the agent while a separate GPU server handles inference and training.
- The four-phase loop runs rollouts, assigns rewards, trains a LoRA on grouped trajectories, then reloads that LoRA into vLLM before resuming; inference is blocked during training.
- Uses LoRA for parameter-efficient fine-tuning and vLLM for multi-GPU inference, supporting most vLLM and HuggingFace transformers compatible causal models, with Gemma 3 noted as unsupported.
Radar summary, generated from the project's public sources
star history
- PR#17 7k 2025-09-17
- now 11k + 3k since first covered
curve is sampled from GitHub's star history, plus our own daily readings since we covered it; the dashed stretch is before we first covered it, the solid line since. figures at coverage are the numbers we printed then (approx.), current count is live.
understory
Output and attention are roughly in balance.
- output, commits & releases
- clout, star velocity
output = commits & releases; clout = star velocity, both 0 to 100 monthly indices; the gap where output runs above clout is the understory. The understory →
covered in
-
Automated Reasoning Trainer
similar projects
compare these →- ⚡ Unsloth
6.8× the stars
Local UI to run and train LLMs and diffusion models, including Qwen3.8, Kimi K3, MiniMax-H3, Gemma 4, DeepSeek-V4, FLUX and more.
72k ACTIVE - 🧠 MaxKB
2.1× the stars · GPL-3.0 licensed
🔥 MaxKB is an open-source platform for building enterprise-grade agents.
23k ACTIVE - 🛠️ Agent Lightning
The absolute trainer to light up AI agents.
17k ACTIVE
comments
Sign in with GitHub to add your blip on ART.