Insight · October 10, 2026
Change the model,
not its notes.
Fine-tuning trains a model on your own examples until a behavior is set into the weights. What it is, how it differs from a prompt, and when it earns its cost.
01 · In one paragraph
What fine-tuning actually is.
Fine-tuning is training a pre-trained model further on your own labeled examples so that a behavior is set into its weights. It changes how the model acts, its tone, its format, its grip on a narrow task, rather than what facts it holds. You reach for it when a well written prompt keeps falling short on a task the model should be able to do every time.
The model that ships from a lab is general. Fine-tuning takes that general model and nudges it toward your specific job. Google frames the move plainly: you take a general-purpose model and adapt it to your domain by training it on a curated dataset of your own examples. The knowledge of how to do the task now lives inside the model, not in the prompt you send.
02 · Behavior, not facts
It edits the weights, so it edits the reflex.
A model is a large set of numbers called weights. Those weights decide how it turns an input into an output. Fine-tuning runs more training on top of the finished model, showing it example after example of the input and the output you want, and lets the weights shift a little toward that pattern. Nothing is bolted on. The model itself comes out slightly different.
That is why fine-tuning changes behavior and not knowledge. It does not hand the model a fact to read. It teaches a way of responding, so the behavior holds without being described again in every prompt. The cost is that the behavior is now baked in. A fact set into the weights cannot be edited later the way a line in a document can, which is exactly why fine-tuning is the wrong tool for a fact that will change next quarter.
This is the cleanest way to tell fine-tuning apart from retrieval augmented generation. Retrieval keeps your knowledge outside the model and feeds it in at answer time. Fine-tuning folds a behavior inside the model before any question is asked. One decides what the model can look up. The other decides how the model acts on its own.
The one line to keep
“A prompt tells the model what to do this time. Fine-tuning changes what it does by default.”
03 · The right jobs
It shines where a prompt keeps slipping.
The behaviors fine-tuning is good at are the ones a prompt never quite pins down. A strict output shape the model drifts away from after a few turns. A house voice that a style instruction only ever approximates. A narrow classification the base model fumbles because it has met the task too rarely. In each case you are not giving the model something to read. You are teaching it a reflex.
Google lists the pattern that keeps coming up across hundreds of organizations: fine-tuning to perform specialized tasks, to adapt to an industry or function, and to improve output formats and styles that match an organization. The named use cases include summarization, classification across text, image, and audio, document extraction, and improving the helpfulness of a retrieval system output. The thread through all of them is a well defined task with examples of the right answer, not a store of facts the model has to recall.
04 · One measured example
A moderation model that went from decent to reliable.
Anthropic published a concrete result when it opened fine-tuning for Claude 3 Haiku in Amazon Bedrock. A team fine-tuned the model to moderate online comments, flagging insults, threats, and explicit content. Fine-tuning raised classification accuracy from 81.5% to 99.6%, and it cut tokens per query by 85% at the same time.
Both numbers matter, and they matter in different ways. The jump in accuracy is the behavior improving. The drop in tokens is the cost falling, because the trained model no longer needs a long instruction and a pile of examples stuffed into every prompt. The work moved from the prompt into the weights. You teach the task once in training rather than describing it again on every call.
Before fine-tuning · 81.5% accuracy · long prompt every call After fine-tuning · 99.6% accuracy · 85% fewer tokens per query The task moved from the prompt into the weights.
Source · Anthropic, fine-tuning Claude 3 Haiku in Amazon Bedrock
05 · How it is done
Four ways to train, from heaviest to lightest.
Full fine-tuning
Every weight is open to change
The whole model is trained on your examples. The most reach, the most compute, and a full copy of the model to store and serve.
LoRA and the PEFT family
Freeze the model, train a small add-on
Google notes LoRA freezes the pre-trained weights and injects smaller trainable layers, so you train a fraction of the parameters and keep the base model shared.
Preference tuning
Learn from chosen over rejected
The model is shaped by pairs that mark a better answer against a worse one, used to align tone and judgment rather than teach a single correct output.
Distillation
A small model copies a larger one
A smaller model is trained on a stronger model output, so a cheaper model carries the behavior of an expensive one on a narrow task.
The lighter methods exist because full fine-tuning is expensive to run and to store. LoRA and the wider parameter efficient family keep the base model frozen and train a small set of new weights beside it, so you own a light add-on rather than a full copy of the model. For most teams that is where fine-tuning starts.
06 · The order that saves money
Reach for it last, not first.
Fine-tuning is the most powerful tool in the set, so teams reach for it first and pay for it twice. Google is direct about the order: in many cases a good prompt and in context learning will be enough, and fine-tuning is where you go as you move toward more specialized production requirements. Start with the cheap move and climb only when it fails.
The ladder has three rungs. First the prompt, which is free and changes in seconds. Then retrieval, when the model lacks facts, which keeps the knowledge searchable and outside the model. Then fine-tuning, when the model lacks the right behavior and no prompt will hold it. Running the training job first, on a problem a better prompt or a retrieval step would have solved, is how budgets disappear.
It is not either or. A mature system often fine-tunes for the house voice and the output shape, then layers retrieval for the facts. For the full decision between the two, see RAG versus fine-tuning.
07 · When not to
A fact that moves does not belong in the weights.
Do not fine-tune to teach the model a fact. A price, a policy, a product detail, anything that changes, is a retrieval problem. Train it into the weights and it is baked in stale the day it moves, and the only fix is another training run. Keep the facts that change outside the model, where editing one line updates the answer at once.
Do not fine-tune before you can measure the result. You need a fixed set of examples and a score, so you can tell whether the trained model is actually better than the base model with a good prompt. Without that, you are spending compute on a change you cannot prove. And do not fine-tune a task you have never gotten working in a plain prompt. If you cannot describe the behavior well enough to prompt it, you cannot build a clean dataset to train it either.
Closing
Teach a reflex, not a fact.
Fine-tuning is how you change what a model does by default. Save it for the behavior a prompt cannot hold, measure it against a plain prompt first, and leave the facts that change outside the model where you can edit them. Name the problem, and the tool picks itself.
Written by · October 10, 2026
Donny Smith · ECD, Founder, Bttr.
Over the past 15 years, he has led creative teams and contributed to products used by millions of people worldwide, working with companies including Apple, Opendoor, JP Morgan, GE Aerospace, Pepsi, and Alterra Mountain Company.
LinkedInShare this perspective
More insights
Adjacent perspectives.
Bttr. Field Brief
The brief Bttr. writes for senior buyers.
Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.