My guess would be that some significant portion of github code published under a permissive license is actually licensed improperly.

Working that out at scale seems intractable, but maybe the training set doesn't need to be as big as I'm assuming.