Atp post-training is much more influential towards model behavior than pre-training data.