All datasets are in tfrecords format. Dataset #2 and #3 each contain manifest.txt files which explain the shape of each datafield.
All data is provided under the Creative Commons BY 4.0 license.