We want you to be able to train a model when you do not have annotations on all the datasets.
A powerful approach for training such a model on a large amount of data without requiring a large amount of labels is to use semi-supervised learning(SSL) technique. SSL mitigates the requirement for labeled data by providing a means of leveraging unlabeled data. Since unlabeled data can often be obtained with minimal human labor, any performance boost conferred by SSL often comes with low cost.
Hence to win this challenge we invite you to read
https://amitness.com/2020/03/fixmatch-semi-supervised/
https://arxiv.org/abs/2001.07685
https://arxiv.org/abs/1904.12848
https://arxiv.org/abs/1905.02249
....