SkyRL
A modular, full-stack reinforcement learning library for training large language model and vision-language model agents, combining a training framework, a Tinker API backend, an agent layer for long-horizon tasks, and a gymnasium of tool-use environments.
SkyRL is a full-stack library for training LLM and VLM agents with reinforcement learning on a user's own hardware. Its skyrl package unifies a modular, performant RL training framework (skyrl-train) with a cross-platform backend implementing the Tinker API for training and inference (skyrl-tx).
A separate agent layer, skyrl-agent, trains long-horizon, real-world agents on multi-turn tool use, and skyrl-gym provides a library of math, coding, search and SQL tool-use environments implemented in the standard Gymnasium API, so a new task can be integrated through a simple interface. The project also integrates with Harbor for training terminal-use agents.