Post #3959272

I'm surprised (not) to see that even the few models trained solely with copyleft-compatible data still don't bother tracking which tokens correspond to which files for attribution purposes - at most they're tacked in after the model is trained already.