Welcome RL Environments to the hub

Hugging Face Hub now hosts reinforcement learning environments in a dedicated section. According to the company, an RL environment gives an agent a task, responds to its actions with observations, and scores the outcome; the resulting rewards can measure agent performance during evaluation or provide a learning signal during training. Environments are split into two parts, tasksets and runtimes, and this release focuses on tasksets.

An RL environment on the Hub is a dataset repo that appears under a new RL Environments filter, located at huggingface.co/datasets?other=rl-environment. The Use this dataset button generates the command to run it in a given framework. Hugging Face states there is no new repo type, no registry, and no sign-up. Environments already exist in Harbor, Verifiers, and NVIDIA NeMo Gym.

Four environment frameworks are registered as dataset libraries, and each framework tag adds the framework's icon to the dataset page plus a generated snippet. A dataset can carry more than one framework tag, since tags describe compatibility and compatibility is not exclusive. Adding a tag does not convert files, and each listed framework must support the files in the repository. Hugging Face says adding a tag does not start a job or sandbox; framework runners execute environments locally or on a supported cloud backend.

The release describes why environments were siloed: custom hubs, runtime registries, independent task datasets, or GitHub lists with custom loaders meant publishing for one framework left users of others unable to load it. The blog post says task data lives on the Hub, while runtime configuration and verifier code can live in the repo or the framework.

Example commands are given for Harbor using an oracle agent that runs a reference solution without calling a model, for Verifiers running a model on the same task directories in Docker, for OpenEnv running an agent such as OpenCode with reward and trace output, and for NeMo Gym evaluation and RL training. The post notes a reward of None means no verifier reward was produced and that the error field should be inspected first.

Tagging is done by adding rl-environment plus framework tags to a dataset card's YAML header. Hugging Face says it opened PRs to tag existing environments including BeyondSWE, Terminal-Lego, Harbor-Mix, NatureBench, Reverse-Text-RL, Multi-SWE-RL-Verified, R2E-Gym-Subset-Verified, Scale-SWE-Verified, Workplace Assistant, Structured Outputs, CFBench, and SysBench. Next steps listed are per-config snippets, structural detection for frameworks with strict layouts, custom task UIs, and automatic tagging, which OpenEnv already does on upload.