An environment is represented as a dataset repository tagged rl-environment; these repositories appear in a new RL Environments filter. The release adds no new repository type, separate registry, or sign-up process.
The Hub has registered four framework tags: Harbor, Verifiers, OpenEnv, and NeMo Gym. A tagged dataset page displays the framework icon and a generated command in “Use this dataset.” Repositories can carry more than one framework tag. Hugging Face says task data live on the Hub, while runtime configuration and verifier code may live in the repository or in the framework; execution occurs locally or on a supported cloud backend.
The post lists existing environments being tagged, including Multi-SWE-RL-Verified, with 2,232 of 4,703 rows passing gold-patch validation, and Scale-SWE-Verified, with 17,202 of 20,181 Python issue-resolution tasks described as producing a clean end-to-end reward signal.
The announcement does not establish that an environment tagged for multiple frameworks is automatically portable between them: each framework must support the repository files, and tags do not convert them. It also does not establish cross-framework comparability of evaluation results.