Post-training and Alignment

Introduction to Post-training and Alignment

Post-training continues training after pretraining so a general language model follows instructions and behaves in ways people prefer. Alignment asks whether those learned behaviors match the intended goals and constraints.

Pretraining teaches a model to continue text. Post-training changes that general text predictor into a model that responds to instructions in a preferred way.

These stages optimize different signals, so they must be evaluated separately. Human preference, instruction following, truthfulness, toxicity, and retained task ability can move in different directions.