# OpenPipe/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!

Repository: https://github.com/OpenPipe/ART
Canonical: https://ross.abutalabs.com/products/art
Homepage: https://art.openpipe.ai
Language: Python
License: Apache-2.0
License Family: permissive
Topics: llms, lora, reinforcement-learning, agent, agentic-ai, grpo, rl, qwen, qwen3
Last push: 2026-08-26T21:18:41+00:00

## Health v2 (maintenance only)
Score: 86/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 98, longevity 38
- inputs: {"age_days": 541, "days_push": 7, "days_rel": 19, "gap_med": 0.0, "n_releases_24m": 59}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 10665, forks 982 (observed 2026-08-28T04:10:43.701171+00:00)

## What it is
ART (Agent Reinforcement Trainer) is an open-source Python framework for training multi-step LLM agents using reinforcement learning techniques like GRPO. It provides an ergonomic client-server harness that lets agents learn from experience, with support for models like Qwen, GPT-OSS, and Llama via LoRA fine-tuning.

## Use cases
- train llm agents with reinforcement learning
- improve agent reliability with grpo
- fine-tune qwen or llama for multi-step tasks
- rl training for agentic workflows
- on-the-job training for ai agents
- serverless rl training without gpu setup

## When to choose
- you want to improve an existing LLM agent's reliability through RL
- you need GRPO-based training with minimal infrastructure management
- you want to train open models like Qwen or Llama with LoRA
- you prefer a client-server setup that can train from any Python machine

## When to avoid
- you only need simple supervised fine-tuning or prompt engineering
- you have no reward signal or evaluation environment for your agent
- you need to train non-LLM models or non-agent workloads

## Facets
- artifact type: framework
- maturity: active
- function: llm-training, reinforcement-learning, agent-framework
- domain: reinforcement-learning, large-language-models, machine-learning
- platform: python, cloud
- tags: grpo, lora, rl-training, agentic-ai, qwen, openpipe, wandb, multi-step-agents, ai-agents, gpu, docker

## Member repositories
- OpenPipe/ART (main) score 86

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:10:43.701171+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:18:12.250779+00:00, confidence not recorded.
  - readme: https://github.com/OpenPipe/ART (fetched 2026-08-28T04:10:43.701171+00:00, sha 2a45fb4c0349)
  - homepage: https://art.openpipe.ai (fetched 2026-08-29T08:17:18.630725+00:00, sha 1cc4fb967501)
  - site_page: https://art.openpipe.ai/getting-started/about (fetched 2026-08-29T08:17:18.634629+00:00, sha a6f8342af960)
- Data as of 2026-08-30T08:39:29.467469+00:00.