“The Music-JEPA model is trained in a fully offline setting using paired audio-pianoroll data and does not require environment interaction.”