Show HN: ART – a new open-source RL framework for training agents
a year ago
- #LLM
- #reinforcement-learning
- #open-source
- ART is an open-source reinforcement training library designed to enhance LLM performance in agentic workflows.
- It uses the GRPO reinforcement learning algorithm to train models based on their own experiences.
- Unlike most RL libraries, ART allows execution within existing codebases while handling the RL training complexity in the backend.
- ART's functionality is split into a client (OpenAI-compatible) and a server (runs on GPU machines).
- The client interfaces with your codebase, while the server manages inference and training complexities.
- Training loop involves inference (rollouts, message storage in Trajectories, reward assignment) and training (GRPO, LoRA updates).
- ART supports most vLLM/HuggingFace-transformers compatible models, excluding Gemma 3 for now.
- Currently in alpha, ART welcomes contributions and feedback via Discord or GitHub issues.
- Acknowledges contributions from the open-source RL community and partners for testing.