Post-training
Also called instruction tuning
Everything done after pretraining to turn a raw text predictor into a helpful assistant: example conversations, ratings, rules and safety training.
After pretraining, a model can continue any text but can't hold a useful conversation. Post-training changes that, in several rounds. The model is shown many example conversations of a good assistant at work. It's trained on ratings of its answers, from people and from automated checks. It practises on problems with answers that can be checked, like maths and code. And it's trained against written rules about what to refuse and how to behave.
This stage is where an assistant gets its manners, its format habits and its limits. It's also where side effects come from, like flattering agreement or over-cautious refusals.
For example
The polite tone, the habit of using bullet points, the refusal to give dangerous instructions and the "Great question!" all come from post-training, not from the web text it first learned from.
Put it to use
Where you'll meet post-training in ChatGPT, Gemini and Claude, how to use it there, and the mix-up most people make.
Daily Bytes Plus is in the Android app.