Key Info

Hugging Face is introducing support for reinforcement learning environments on the Hub, arguing they should be handled just like datasets since they consist of tasks, tests, containers, and reward functions.

Highlights

  • RL environments currently live in fragmented places: custom hubs, runtime registries, GitHub lists with custom loaders
  • Publishing for one framework means users of other frameworks can't load it
  • Environments are composed of tasks, tests, containers, and reward functions — making them dataset-like
  • The Hub already versions, gates, and previews data for millions, so no second system is needed