Definition
Post-training is the phase of model development that adapts a pretrained model toward intended tasks and behavior. It starts from a model whose parameters were learned during pretraining. Further training targets particular tasks or behaviors through its data and training setup. Some methods retain a next-token objective, while others use preference or reward signals.
The phase can include supervised fine-tuning on instruction-response examples, preference optimization from comparisons, and training focused on safer behavior. These are separate methods within a post-training pipeline. Post-training is not one algorithm or one training objective.
Simple example
Consider a pretrained language model that can continue text but does not reliably follow user instructions. A model developer first runs supervised fine-tuning on reviewed conversations. The model then learns from comparisons in which reviewers prefer one candidate response over another. Safety-focused examples and evaluations guide another training round before the chat model is released.
Those model updates belong to post-training. Evaluation guides the process but does not itself update the model. The methods can use different data and objectives. Their failure modes differ too.
Why it matters
Two autoregressive model variants can share the same pretrained base and still behave quite differently after post-training. Both continue to predict the next token. The post-trained variant may simply assign much higher probability to responses that follow instructions or use the required tool-call format.
A benchmark result for the pretrained base does not describe every post-trained variant. The post-training recipe can improve target behavior, introduce regressions elsewhere, or teach the model superficial patterns from its training data. Keep the model version tied to the training data provenance and evaluation results.
One important nuance
Post-training is not a synonym for safety alignment. Safety can be one goal, alongside instruction following, reasoning, tool use, or domain behavior. Training for safer responses also does not replace application controls such as authorization and output validation.
In current LLM practice, continued pretraining is usually treated as distinct from post-training. Continued pretraining may happen later in time, but it keeps using a pretraining-style objective on additional data. Post-training typically refers instead to later adaptation toward selected behavior or tasks.