the models don't retain their full training data set