Coffee & Cake: Pretraining, Fine-Tuning and Post-Training

Pretraining, Fine-Tuning, and What Comes After

Ask a language model whether you can move Tuesday's flight, and three versions of the same model may behave very differently. The first continues your question as if it were a line in a travel forum. The second replies like a support assistant and asks for the missing booking details. The third does that too, but is especially careful not to invent a fare rule it cannot know.

It is tempting to place these versions into three tidy boxes labelled pretraining, fine-tuning, and post-training. That picture is simple, memorable, and misleading. The terms do not name three rival techniques arranged on a compulsory conveyor belt. Two describe stages in a model's life. The other describes a way of adapting a model that already exists.

Here is the map to keep. Pretraining is the broad initial training stage. Fine-tuning is a method of adapting an existing checkpoint, meaning a saved state of the model. Post-training is the wider period of work after pretraining. Supervised fine-tuning often happens during post-training, so the very same update can correctly carry both labels. Fine-tuning says something about how the model was adapted. Post-training says where that work sits in its history.

Begin with the airline model before it is an airline model. During broad pretraining, a language model processes an enormous and varied collection of text. At each position it tries to predict what comes next. When its prediction differs from the observed text, that error helps adjust many of the model's numerical settings, usually called weights. Repeated across vast amounts of material, this builds a base model with broad patterns of language and many emerging capabilities.

The base model may know a great deal about airports, dates, refunds, customer conversations, and the usual shape of a support page. But pretraining did not necessarily teach it to act as a dependable assistant. Asked whether Tuesday's flight can be moved, it might write a plausible continuation, imitate a discussion thread, or confidently fill in a policy that was never supplied. Broad capability and useful product behaviour are not the same thing.

Now start from that saved base model and show it curated examples. A customer asks to move a flight. A good response requests the booking reference, explains what can be checked, and avoids claiming that a change has already been made. Another example demonstrates how to handle a missing date. During supervised fine-tuning, the target answer provides the text the model should learn to make more likely. Training adjusts the model from what it already knows rather than beginning again from nothing.

This stage is fine-tuning because an existing checkpoint is being adapted on more specific examples. It is also post-training because it occurs after broad pretraining. There is no contradiction. Think of the two labels as answers to different questions. What method was used? Supervised fine-tuning. When in the lifecycle was it used? During post-training.

Post-training is an umbrella, not one algorithm. After supervised examples, developers might compare several possible answers. They may prefer the response that admits uncertainty over one that invents a baggage rule. They may prefer a concise request for one missing detail over a long refusal. Those comparisons become a learning signal that shifts which kinds of answer the model is likely to produce.

There are different ways to use that signal. One family of methods learns a separate reward model from rankings and then uses reinforcement learning to improve the language model. Another can optimize the language model more directly from preferred and rejected answers. Some systems also use rewards that a program can check, such as whether a constrained answer satisfies a verifier. These can all belong to post-training. No single recipe is required, and not every model passes through every method.

What changes throughout this work is not a little rulebook with a page marked airline manners. The learning signal changes numerical parameters and therefore changes the probabilities of future outputs. A relatively small set of carefully chosen examples can redirect behaviour strongly because it builds on capabilities already formed during pretraining. Small data does not necessarily mean a cosmetic effect.

Nor must every original weight be rewritten. Full fine-tuning can update the whole model. Parameter-efficient approaches can leave the base weights frozen and learn a much smaller set of additional adjustments. When those adjustments are loaded, the effective model behaves differently even though the original checkpoint remains intact. Fine-tuning describes adaptation from a starting model, not one fixed amount of computational surgery.

Return to our airline. After general post-training, one company may adapt the assistant again using examples from its support style and domain. That narrower update may also be called fine-tuning, even though the starting point is already a post-trained assistant. The word tells us that one checkpoint was adapted into another. It does not reveal the model's entire family history.

There is a useful boundary case. A developer may take an existing model and continue the original next-token training objective on a large body of aviation text. That begins from a checkpoint, yet researchers often call it continued pretraining rather than ordinary task fine-tuning. The label matters less than the concrete questions: what data was used, what objective supplied the learning signal, and which parameters changed?

Now separate training from what happens when the product is used. If the airline places today's refund policy inside the prompt, the model's weights have not changed. If the system retrieves that policy from a document before answering, the model has not been fine-tuned. Both mechanisms can improve the current response by changing its input. They do not create a lasting trained update.

The same boundary applies to action. Fine-tuning may make the model better at recognizing when a customer wants to change a booking or at requesting the right details. It does not itself move the flight. A separate tool must contact the reservation system, and product logic must decide what the tool is allowed to do. Training shapes the model's output. Tools and surrounding software turn an authorized request into an external action.

The old slogan that pretraining teaches knowledge, fine-tuning teaches a task, and post-training adds manners therefore falls apart. Pretraining can build broad skills as well as factual associations. Fine-tuning can shape a task, a domain, a format, or a preference, and may sit inside post-training. Post-training can change whether the model attempts a problem, which strategy it follows, and how it handles uncertainty. It is much more than polishing the tone.

None of these stages guarantees truth. The objective is always a proxy for what people actually want. Curated answers can contain mistakes. Human preferences can be inconsistent or reward an attractive shortcut. A model trained to satisfy an automatic checker may learn to satisfy the checker without solving the deeper problem. Improving one target behaviour can also weaken another or make the model too specialized for its training examples.

Terminology varies as well. Laboratories draw the borders differently, especially around continued pretraining and the many forms of preference learning. That is not a reason to abandon the terms. It is a reason to ask for the mechanism beneath them instead of treating the label as a complete technical description.

So when somebody says a model was tuned, ask four questions. Which checkpoint did the work start from? What supplied the learning signal: broad text, curated answers, comparisons, or checkable rewards? Which weights or added parameters changed? And what behaviour was tested afterward, including what may have become worse?

Those questions restore the map. Pretraining builds the broad base. Fine-tuning adapts a base that already exists. Post-training names the larger workshop after pretraining, and supervised fine-tuning is often one of the tools used inside it. The terms overlap because they describe different dimensions. Once that is clear, the lifecycle stops looking like three boxes and starts looking like what it is: a sequence of choices about where a model begins, what signal changes it, and what kind of behaviour should emerge.