Pretraining teaches the model to continue text. Post-training teaches it what kind of continuation is acceptable, useful, formatted, safe, and rewarded. This is where a base model becomes the assistant people actually interact with.
Much of what feels like “assistant personality” — refusal behavior, formatting discipline, tool-use style, conversational helpfulness — comes from post-training and system scaffolding, not from pretraining. Users experience post-trained assistants, not raw base models.
The post-training pipeline is a sequence of stages, not a single step: continued pretraining, supervised fine-tuning (SFT), preference training (RLHF, DPO), and reinforcement learning with verifiable rewards. Each stage shapes different aspects of model behavior, and critically, they differ in which tokens count toward the loss — they are not merely “same training, different data.”