[lilyball]

One pretty useful reason would be to then eliminate that content from subsequent training data, so you're not training the next model on the previous model's output.