Fine-tuning: teaching habits, not facts
The pricing page said train a model on your data, and for months I assumed that meant the model had learned our documents - that a fine-tuned model would know our policies the way it knows arithmetic. It does not, and one sentence is the whole correction: a fine-tune is an extra training run over examples you supply, and what it reliably installs is habits - a tone, a format, a way of moving through a task - not facts you can update. The facts problem is mechanical. Training changes the weights, the numbers inside the model, and anything written into them is frozen at the moment of the run, so when the underlying source changes you do not have a current model; you have a stale one that sounds confident. That is why the working order is: prompt first, because it is minutes and reversible with an edit; retrieve second, showing the current document at question time whenever the facts move; fine-tune last, when behavior is what is broken - the model knows the material and keeps answering in a shape your product cannot use, no matter how the prompt says it. When people want the training job cheaper, the common shape is an adapter: the usual one arrives under the name LoRA, which freezes the base model and trains a small add-on that plugs in beside it, so the run costs a fraction of a full retrain and the general abilities stay intact. And distillation in a line: have the big model write thousands of good answers, train a small model on them, and keep most of the quality at a fraction of the serving cost. The sentence that earns the seat in the budget meeting is the question, not the yes: how often do the facts change, and who owns them.
Most people have this backwards
Training a model on your own documents usually makes its answers about those documents worse over time, not better.
The catch
It holds whenever the documents change, because the run froze a snapshot and every later edit leaves the model confidently citing the old one, while the base model plus retrieval always reads the current version. It weakens for material that barely moves - a house style, a fixed format, a domain motion - where the tuned model genuinely beats a prompt, since habit in weights outlasts any instruction. The question that locates the line: what share of your training examples will be wrong a quarter after the run?
Baseline first - your honest read.
A teammate proposes training a model on the company wiki so it knows everything internal. Which sentence is your answer in the meeting?
Make each call, then see the verdict
Predict first
A model is fine-tuned on the current product docs, and the docs change weekly. Over one year, does the share of its answers that are factually current rise, fall, or stay flat?
Four jobs are asking for a training run. Pick the lever each one actually needs.
Drafts must come out in a fixed house voice - short, no headers, a specific sign-off pattern - and prompt rules keep losing the sign-off every time the prompt grows.
Answers about pricing keep coming out wrong, the pricing page changes quarterly, and someone proposes a tuning run on the current page.
An email summarizer works fine. A teammate proposes tuning a model on past summaries to make it better at representing us.
The feature works and costs too much per answer. The team has heard a smaller model could carry it and has heard the word distillation.
How it works
Each lever touches a different layer, which is why the decision is a sorting problem and not a menu. Prompting changes only the text in front of untouched weights: instant, reversible, and unable to change the model itself, so its fixes last exactly as long as the text carrying them. Retrieval changes what the model can see at question time: the right fix for moving facts, since an index refresh is cheap and a retrain is not, and it leaves behavior untouched. A fine-tune changes the weights through demonstration: you author the inputs and the exact outputs you wish the model produced, the run adjusts the numbers until those outputs become likely, and the model comes back carrying whatever the examples collectively demonstrated - the format, the tone, the shortcuts, including ones nobody meant to teach. That last sentence is why every tuned version ships behind a graded set: tuning moves everything it statistically can, and nobody reads every behavior change in advance. The adapter route changes the economics without changing the discipline: freeze the base, train the small add-on, and a tuning job becomes cheap enough to repeat - adapters can be swapped in and out of one base model, so one tuned voice can coexist with another on the same serving bill. What stays expensive is everything around the run: the examples to author and keep honest, the eval run against the previous version, and the version to roll back to. Order them prompt, retrieve, tune, distill-for-cost, and most fine-tuning budgets spend themselves correctly.
The house band
The picture is a ten-year session musician joining a house band. The player's general ability is the base model, and nobody rehearses it - it is already there, and drilling scales would add nothing. What the band rehearses is the arrangements: this tempo, this ending, the cue that says the last chorus arrives early. That is a fine-tune - habit, installed through repetition of a specific shape. The sheet music on the stand is retrieval: nobody memorizes a new chart each week, they read the current one at the gig, and when the chart is revised the stand gets a new page instead of a decade of muscle memory being retrained. And the adapter shortcut has its own picture in this band: a session player who keeps a private technique and learns only this band's arrangements, which is cheap compared with rebuilding the player.
Where it breaks
The analogy breaks at where the repetition lands. A musician rehearses one arrangement and remembers it separately, while a training run is one pass that nudges everything the examples statistically touch - the rehearsal meant for the sign-off pattern quietly rewrites unrelated habits, which is the single sentence behind why every tuned version ships behind a graded set. And the band forgets nothing on schedule: weights keep every fact they were trained on until a later run replaces it, so there is no arrangement the model quietly leaves behind.
Why interviewers lean on it
Fine-tuning questions are budget questions wearing a technical costume, which is why they arrive early in interviews. A tune touches four budgets at once: authoring the examples, the compute of the run, the eval that proves the version, and the maintenance that starts the day it ships. The follow-ups press exactly where the line sits: name the one behavior this run improves, say what share of the training examples will be stale a quarter later, and describe the rollback when the graded set moves the wrong way. A candidate who has run one answers with the adapter economics - the small add-on instead of the full rewrite, the frozen base keeping general ability - and with distillation as the cost play behind an eval gate. A candidate who has not answers with the marketing vocabulary: custom, trained, learned. The product half is the same bill: tuning a moving fact set is a subscription to retrains, and the engineer who recommends it hands the company a recurring cost nobody saw itemized. The strongest position in the room is not that fine-tuning is bad; it is the two questions - do the facts move, and is behavior the thing that is broken - because that is the sound of someone sorting levers instead of shopping.
The follow-up chain
Interviewers do not stop at the surface answer. Answer the way you would out loud, or skip - skipping reveals the strong answer and ends the chain.
The interviewer's next question
You keep calling fine-tunes habit training. Prove it: what exactly goes in your training set, and what do you refuse to put in it?
Now answer one you have not seen
A teammate proposes fine-tuning on three years of past support tickets and retiring both the prompt library and the retrieval stack that internal tools use today. What do you push back on, what survives, and what would you measure first?
Expires fast - as of 2026-10-06
Adapter fine-tuning, usually under the LoRA name, is the default way small teams fine-tune as of 2026: the base stays frozen, the add-on trains cheaply, and hosted platforms expose it as a checkbox-level feature, which is why this concept treats the adapter as the common route and the full retrain as the exception.
Distillation for serving cost is the most-cited fine-tune use case in 2026 practitioner writeups, ahead of voice and format tuning, and the graded-set gate before routing traffic is the standard caution paired with it.
The phrase train a model on your data appears across hosted-platform marketing in 2026 without saying what survives a later document change; this concept stays on training mechanics rather than any platform's wording.