Welcome RL Environments to the hub

In a post dated September 28, 2026, Hugging Face says reinforcement learning environments can live on its Hub as ordinary dataset repositories carrying an rl-environment tag. An RL environment is a task for an agent: it presents a situation, returns what the agent observes after each action, and scores the outcome so the agent can be trained or evaluated. The post says there is no new repository type, registry or sign-up.

The point is reuse. The Hub handles hosting, discovery and versioning of the task data, while a framework executes the environment, locally or on a supported cloud backend. Four frameworks are registered as tags, Harbor, Verifiers, OpenEnv and NeMo Gym. A tag adds an icon and a generated loading snippet, and does not start a job.

This release covers task sets, not runtimes. The post's examples include running a task's reference solution without calling a model, which checks the task itself, and a rollout that prints a verifier reward. It reports opening 12 pull requests to tag existing datasets, which is not evidence of broader adoption.