Post-training data requires strict stylistic consistency across different subject domains to prevent the trained model from exhibiting irregular formatting behaviors