ART: Minimal Code Changes, Maximal LLM Performance Gains via RL

Agent Reinforcement Trainer (ART) is an open-source reinforcement learning library designed to boost Large Language Model (LLM) performance in agent workflows. Leveraging the powerful GRPO algorithm, ART trains models from their own experiences. Unlike most RL libraries, ART integrates seamlessly into existing codebases, offloading the complexity of the RL training loop to its backend. ART consists of a client (for interacting with your code) and a server (handling inference and training). The training loop involves inference (gathering data and assigning rewards) and training (using GRPO to train the model and update LoRAs). ART supports most vLLM/HuggingFace Transformers compatible causal language models. Currently in alpha, ART welcomes contributions.