rLLM: Framework for Reinforcement Learning of LLMs
What it is
rLLM is an open-source framework that enables training language agents using reinforcement learning by integrating any agent harness with various training backends. It standardizes the pipeline of running agents, collecting traces, computing rewards, and updating model weights, allowing the same agent code to be used for both evaluation and training.
The problem it solves
The fragmentation between agent development harnesses and RL training backends, which previously required significant code refactoring to adapt agents for reinforcement learning.
What you could use it for
- Training LLMs on agentic tasks like coding or math using existing agent frameworks
- Evaluating language agents on 60+ integrated benchmarks such as SWE-bench and GSM8K
- Switching between distributed (verl) and single-machine (tinker) RL training backends via CLI flags
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.