Author here!

This is a proof of concept for a tiny word -> palette model.

I extracted data from all kinds of sources, and I plan to do more (Wikidata ingredient lists so I can color stuff like "burger" and "guacamole"), and possibly an image color extraction later (would make extraction significantly more expensive).

That data extraction is the only part of the pipeline that is not open (data is commited though). I plan to open it in the future as a separate project, since it does more than extracting for this particular PoC.