Donating training data

[haydn-jones]

I'm planning on generating some amount of training data (hopefully on the order of 10s of thousands of pages, depends on cost) using olmocr/data/buildsilver.py (at least I assume this is how the data was generated). I've been running this on a lot of medicinal chemistry-esque papers and its been struggling here.

If you are open to me donating the data, I can use open access papers exclusively for this, otherwise I'll just throw what I have into a private mix.

[jakep-allenai]

I've asked the team, definitely would be some legal considerations. Would be nice to have some more chem data, and the buildsilver script is the way we did the original olmocr-mix-0225. I'm also trying some new things out on the jakep/new_data branch, but haven't had better results with other models yet.