When we are training a model, we already have the ground truth which the model will use to improve on.

Labeled data, and mapping inputs to outputs.