Configuration "preload_from_hub" does not support "repo_type" (preload datasets)?

[cboettig]

Bug description.

Spaces configuration supports the field preload_from_hub according to the documentation. The contents of this field get passed to huggingface-cli download, which defaults to repo-type="model". It appears there is no way to configure this to pre-load a "dataset" repo type instead, despite the documentation description:

This is particularly useful for Spaces that rely on large models or datasets that would otherwise need to be downloaded at runtime.

Describe the expected behaviour

Provide (or document) a method to pre-load datasets. The examples all show model-repo type.

Additional information

I know we can work around this by committing data to a space, or downloading, or using a model repo instead of a dataset repo to store the data, but these are clearly not ideal solutions.

[julien-c]

maybe cc @apolinario given #1156