MagnaTAgaTune split that I've used for my experiments in the papers
Sample-level Deep Convolutional Neural Networks for Music Auto-tagging using Raw Waveforms, SMC(to be published), 2017 and
Multi-Level and Multi-Scale Feature Aggregation Using Pre-trained Convolutional Neural Networks for Music Auto-tagging, IEEE SPL(to be published), 2017.
We filtered out the tags and used most frequently labeled 50 tags and left a list with more than 1 tag out of 50 tags. Also, only songs with more than 29.1 seconds of audio were left.