Skip to index

GLOSSARY

Training (Pretraining & Post-Training)

The process that creates a model: pretraining learns language from oceans of text, post-training shapes it into a helpful assistant.

Pretraining is the expensive part: months on GPU clusters, learning next-token prediction from web-scale text, code and (lately) images. What emerges is a raw “base model” — brilliant at continuation, useless as a product. Post-training then transforms it: supervised fine-tuning on curated dialogues, RLHF or DPO on human preferences, and specialized passes for reasoning, coding or safety. This is why a “small” post-trained model can outperform a larger raw one.

Almost nobody trains frontier base models — that's a five-company game — but post-training is democratizing fast: open models ship with recipes, and fine-tuning a released checkpoint costs hundreds of dollars, not millions. When a vendor touts “custom models for you”, it almost always means post-training.

Related terms

Tools that use this

Related categories