automated combined public dataset

[mschrimpf]

we currently have a set of public datasets, either as splits or entirely private.
It could be quite impactful for the community to automatically and periodically combine these into a large dataset release, e.g. for model training.