I think the difference is when something is actually output and then removed after you already see it... that doesn't seem to be a training data issue