Really wish they published the whole dataset. They don't specify on the page or in the paper what the full set was like, and the GitHub repo only has one of the easy-to-read ones. If >=10% of the set is comprised of clock faces designed not to be readable then fair enough.