Any progress? Does fine tuning add pdf knowledge in to the model? Working any better than prompt based methods?