Training data that provides detailed annotations for every part of an input, rather than just overall labels.