That doesnt make training the models wrong, just using those specific outputs wrong.