Hugging Face introduced reinforcement learning environment support to its Hub on September 28, 2026, indexing task datasets under a dedicated filter rather than creating a separate repository system. The release focuses on tasksets, allowing existing dataset repositories to host environment files while external frameworks execute them locally or on cloud infrastructure.
Hugging Face confirmed that repositories tagged with rl-environment appear in the new directory. The platform does not require separate user accounts or new repository formats. Datasets can carry multiple framework tags, which generate interface snippets showing the required installation and execution commands for each tool.
Framework support
Four environment frameworks are registered as dataset libraries: Harbor, Verifiers, OpenEnv, and Nvidia NeMo Gym. Harbor executes task reference solutions using an oracle agent to inspect verifier output without querying a model. The Harbor integration within Verifiers runs those task directories across runtimes such as Docker using execution harnesses that include bash.
OpenEnv runs identical task directories with agents such as OpenCode, returning rewards alongside agent traces. The setup routes connections through a temporary Gradio tunnel so sandboxed agents can reach the OpenEnv model proxy. Nvidia NeMo Gym gathers trajectories and scores rewards for evaluation and training, including schema adherence checks within its Structured Outputs dataset.
Tagged benchmarks
Hugging Face opened pull requests to tag existing evaluation sets across the community. Tagged repositories include Scale-SWE-Verified, which covers 17,202 verified Python issue-resolving tasks from a total of 20,181, and Multi-SWE-RL-Verified, which retains 2,232 of 4,703 rows validated across C, Go, Java, JavaScript, Rust, and TypeScript.
Other tagged datasets include Harbor-Mix, which contains 100 tasks selected from the Harbor adapters pool, the 90-task NatureBench, and Workplace Assistant, which includes five databases, 26 tools, and 690 business tasks.
Hugging Face plans to add per-configuration command snippets, layout detection for strict directory structures, and broader automated tagging across framework upload paths. OpenEnv already applies the compatibility tags automatically during dataset uploads.
