When creating a training dataset, what is the recommended number of samples for the Classification fields?