Or better yet: you include it, and actually evaluate it as part of the mix of expected clients. Deliberately excluding certain classes of datapoints doesn't strike me as something that'll produce a maximally-useful dataset.