Full announcement & details 👉 openpipe.ai/blog/art-tra...

Join us in shaping ART:
- Discord: discord.com/invite/F6Mxp...
- GitHub: github.com/openpipe/art

✨ Key Innovations in ART-chitecture:

- Decoupled frontend (user logic) and backend (inference/training).
- VRAM optimization, enabling training 7B models even on free-tier Colab!
- Builds on
@vllm_project
, TRL by
@huggingface
and
@UnslothAI

3️⃣ Seamless integration: Other RL trainers require substantial refactoring to fit existing codebases. ART is designed for plug-and-play compatibility, easing integration with tools like CrewAI and the OpenAI Agents SDK.

2️⃣ GPU efficiency: Typical RL rollouts can leave GPUs idle waiting for external tasks. ART separates frontend (rollouts, reward logic) and backend (inference, training), allowing parallelized execution and higher GPU utilization.

What makes ART different?

1️⃣ Multi-turn roll-outs: Existing RL frameworks often handle single-turn interactions. But real-world agent tasks—like web navigation—are multi-turn. ART natively supports multi-turn agent rollouts, essential for real-world agentic flows.

🚀 Early Results: we’ve included example code for the following models (more details soon)
1. Generating HN titles (surpasses all SOTA)
2. 🕹️ Tic-tac-toe (7B model surpassing GPT-4o)
3. 🔍 Clue (14B model beating frontier models)

All with runnable examples in the repo!