The MUSDB18 dataset is a dataset of 150 full lengths music tracks (~10h duration) of different genres along with their isolated drums, bass, vocals and others stems. The website also provides information on the "Compressed STEMS" format used in the dataset.
The Open Source Tools & Data for Music Source Separation site, by Ethan Manilow, Prem Seetharaman, and Justin Salamon, provides an in-depth tutorial on source separation, and covers the main steps in source separation, including importing datasets and training models.
The Deep learning for automatic mixing book, by Christian J. Steinmetz, Soumya Sai Vanka, Marco Martínez and Gary Bromham, provides a comprehensive tutorial that serves as a starting point for researchers interested on applying the last deep learning techniques on automatic music mixing. The video of the tutorial is available here.