This brings up the interesting question of whether it's ok to publish the trained model and not the data?

I'd imagine it's sufficiently derivative as to be allowed? What if you wrote a model that produced summaries of the book - would that be allowed?