principles.fyi · the brain · concept

post-training

All the training that happens after pretraining to turn a raw model into a usable assistant.

pretrained base -> instruction tuning + preference alignment -> aligned assistant

Post-training is everything that comes after the giant pretraining run: it takes a model that merely continues text and shapes it into one that is helpful, honest, and harmless. It bundles instruction tuning (teaching it to follow requests) and preference alignment (teaching it which answers people actually prefer), and is loosely stretched to cover test-time-compute tricks like chain-of-thought reasoning. Almost none of the model's raw knowledge comes from here — that was laid down in pretraining; post-training mostly steers behaviour rather than adding facts.