If you train a model on your filtered data and then use that model to filter the data you'll train on... it might become impossible to know what your data actually represents.