I like this dump of Discogs.com metadata

Using data mined from Discogs.com, connections can be drawn between regional music scenes and their sometimes brief and ephemeral artifacts. Unfortunately, the metadata in this dataset is not organized or structured very well, which limits interoperability. There are no barriers to accessing the data, and in fact Discogs.com has an open API so that users can pull data directly from their site. This would give researchers the opportunity to structure the data in more useful ways.

I'm also fascinated by this MetaMapper, a "data discovery and documentation platform for improving how teams understand and interact with their data." Data science can be greatly enhanced by well documented, discoverable, and re-useable data. This tool looks useful for these purposes by automating the process of documenting data - a time-consuming and tedious task.

And this Lesson guide and textbook for "Data as a Science" course. Any source that helps with grappling data science is welcome in my book!