Great work! Could you please share the GPU hours used for training the draft model, as well as for re-annotating the dataset?