Replicating dataset splits?

[Daniel Stoller]

I want to compare the results in this paper on the three datasets to my own approach. How do I make sure that I have the exact same data including the same training/validation/test split?

For example for the Blizzard dataset, i couldn't see an "official" split, so I suppose you used your own one for the paper? In that case, where can I find details on it so I can split it the same way?

Similarly for the Piano dataset, I am wondering how to get the dataset in full format (full song files, not 8-sec chunks) and partition it in the same way you guys have done it.

Otherwise it's very hard to compare the results to any other method.