Given a test dataset from a different distribution from train, filter out examples that are too 'novel'.
