Thanks for your great work. I am trying to replicate the result, may I ask for the exact train/valid split of the SongEval dataset in your paper.