Training with diverse types of labeled data from different sources to improve generalization across varied scenarios.