I'd rather vaguely gesture at data in the training corpus. That, at least, introduces some variance and forces me to add my spin on things after testing.