Even if they don't release the (copyrighted) training data itself, do they provide a "recipe" to reproduce it?

Like, a list of sources used and how they were harvested?