And fine tuning datasets will be compressed and sent rather than the whole model