huggingface_hub
huggingface_hub copied to clipboard
Adding a new "model tag category": Environments for Reinforcement Learning task
Hey team,
So I'm thinking about how we could add a new category for environments (a little bit like datasets).
A little bit like this (just a crude illustration):
The thing is should we list environments based on the environment lib or the environment itself? Wdyt @thomwolf, @ClementRomac, @osanseviero and Edwards?
For instance gym have like:
- Pong
- Space Invaders
- Cartpole
- MountainCar
- etc etc
Maybe the idea is to use category for the env lib like:
- gym-classic-control
- gym-atari
- vizdoom
- ML-Agents
- virtual-house
Yes, nice!
I think we should have probably have both framework and environment name indicated given that some environments are available through different frameworks (with usually subtil differences), e.g. atari-py
or ale-py
(but I think there are other case of similar envs accessible through different frameworks)
We should maybe even have version numbers for environments and framework in the metadata as much as possible. Doing a good job in terms of reproducibility and discoverability is key in RL and this applies to environments as well as algorithms
On another note, the spaces associated to an environment could probably be an editor or explore-interface of the env when it's possible, wdyt?
Also cc @gary149 @julien-c since this is a product change
no informed opinion on whether we want to link to envs, or env libraries (or both). Maybe something like {library}-{env}
would work well?
As for the metadata format, we could either:
- add a
environments:
top-level yaml key to model card metadata, like we have to link todatasets:
- or, use
tags:
– like we do for library names – which has the drawback that we need an exhaustive list of potential env names
cc @osanseviero and @Pierrci on the metadata format
(note that we would need a list of potential env names to map them to documentation pages, anyways – if we do want to link to their doc page as in your mockup)
I think an environments:
top-level key would work well yes.