For the first time, we also added a mid-training stage before RL. We trained on several trillion tokens of code, scientific articles, blogs, and more, saving the scarcest, highest-quality data for last.