When supervised fine-tuning is the wrong first step
Supervised fine-tuning is the wrong first step when nobody can write one good answer, when two people mark the same input differently, or when the real gap is looking up text that changes. The method itself is ordinary: labelled input-output pairs, a train slice, an unused slice, and next-token loss on the targets. That is a behaviour job. It is not a way to settle a tone argument, invent a schema, or skip permission. If those artefacts are missing, the honest next step is readiness, retrieval, or a written workflow — not a training booking.
By Shen Pandi, Towards AGI. Reviewed September 2026. This page answers: when supervised fine-tuning is the wrong first step.
Who this page is for
Open this URL when the plan already says “we will fine-tune” and the folder in front of you is still just documents. The usual readers are a handler who will have to live with whatever the labels say, a data owner who has not cleared the files for training, and a sponsor who wants a visible GPU week before anyone writes a schema. This is not a trainer walkthrough. It will not score SFT against LoRA, DPO, or a vendor chat style on your book.
If the artefact you care about is a small adapter file, read LoRA after pairs exist. If someone wants to rank two already-correct drafts, read preference alignment. If facts move every week, the commercial fine-tuning page is where retrieval sits. This page owns labelled pairs and the moments you must not train yet.
A labelled pair is the only training object
An SFT example is one input the model will see in production, one target completion a named person will defend, and a written refuse rule for the cases that should not complete. That can be an extract with a source quote, a queue label that already exists on the desk, or a letter a handler would actually send. A pile of PDFs is not a pair. A sharepoint dump is not a pair. “The model should know our house style” is not a pair.
Write the schema before anyone opens a trainer. Name the fields, the allowed values, the exception token, and the person who wins when two marks collide. If that sentence cannot be written, you do not have supervised data. You have a pile. Training on a pile copies the pile’s accidents: one author’s voice, a stale template, a silent skip nobody noticed, a tone that half the desk already rejects.
Production-shaped input, not a cleaned demo paragraph.
Target completion a named specialist will defend.
Refuse or exception behaviour in the same schema.
Owner who resolves two colliding marks on one input.
When you must not train yet
Do not train if the files are not cleared for training and evaluation. Permission is not a modelling problem. Do not train if two specialists mark the same input differently and nobody owns the tie — especially when the fight is tone. The optimiser will copy the louder annotator at GPU prices. Do not train if the answer’s facts change every week and nobody proposed retrieval or a process change. Last month’s handbook baked into weights is still last month’s handbook.
Do not train if there is no unused comparison pack that can stay out of the optimiser until the end. Tidy training files are the easiest way to fool a steering group. Do not train if nobody owns the moment the model should refuse. A model that always completes is not helpful on a regulated letter. It is an unowned exception path wearing a completion.
No written definition of a good answer.
No unused comparison pack that can stay unseen until scoring.
The gap is lookup of changing text, not stable behaviour.
Nobody owns review when the model should refuse.
Tone fights are not a dataset
The most common insurance stop we see is not a missing GPU. It is two handlers who already disagree about how a correct letter should sound. One wants clipped and formal. The other wants warm and explanatory. Both drafts can be factually aligned with the same case. There is still no single target. Booking SFT now teaches whichever voice landed in the spreadsheet that week.
Settle the letter standard first. Write the tone rule a conduct reviewer would sign. Then, and only then, can a pair exist. If the remaining argument is still taste after that rule is written, you still do not have supervised data. You have a culture problem. Culture problems do not get cheaper when you add an optimiser.
What SFT actually changes
SFT shows the model pairs and nudges weights or adapters so later completions look more like those pairs. That helps terminology, format, restraint, and a repeatable workflow. It does not add a fact that was never in the context. It does not move prompts into your building. It does not convert a community-licensed base into Apache. It does not make a reserved-test number appear because train loss went down.
People say “fine-tune” for SFT, for preference methods, and for a vendor UI that stores a style on their stack. Write which one you mean. This page is the supervised form: one target per input, next-token (or equivalent) loss on that target. Ranking two otherwise-correct drafts is a later artefact. See preference alignment after a stable SFT baseline exists.
What changes in the run
You cut labelled pairs into a train slice the optimiser may see and an unused slice it may not. Validation, if you use it, is only for choosing settings. Loss is next-token prediction on the target tokens — or the same idea under a chat template. After the run you hold updated weights or adapters that complete more like the pairs on similar inputs: format, terminology, the exception token you taught. You do not hold a new foundation model.
The licence on the checkpoint is unchanged. Inference location is unchanged. If two specialists still disagree on a label, the run copied the disagreement. If someone leaked the unused pack into training because they “just wanted to check”, you no longer have an independent comparison. The data-readiness review is the engagement that writes those splits down. This page assumes that work is done, or that you are about to discover it is not.
Insurance: letter tone still in dispute
Handlers disagree on how a correct letter should sound
insurance
A claims desk wants suggested correspondence for one document family under handler review. Two senior handlers already fight about tone on letters they both accept as factually aligned: one marks the warmer draft as unprofessional, the other marks the clipped draft as rude. There is no written tone rule. Booking SFT now is the wrong first step. The run will freeze whichever voice won the last meeting.
The first work is a one-page letter standard a conduct reviewer will sign, a tie-break owner, permission to use the files, and unused test documents the optimiser never sees. Until those exist, do not train. Authority for what leaves the building stays with the people you already named. The model may later suggest a letter. It does not settle the claim and it does not settle taste.
Retail: attributes once labels already agree
Agreed attribute labels, then unused SKUs, then SFT
retail
A catalogue team already has an allowed-value list merchandisers will actually use: colour, material, fit, and a missing-fibre exception. They can show pairs that follow that list. SFT becomes the conversation only after unused test documents exist — a small reserved set of SKUs the optimiser will not see, judged by the same merchandisers. Agreed labels without that reserved set are still a way to fool yourselves on tidy training rows.
Once the unused pack exists, SFT can teach format and restraint: fill the fields, flag missing fibre, do not invent a care code. If supplier copy is the system of record and changes nightly, retrieval or a feed is the knowledge path. SFT on last season’s dump will not track this week’s intake. Do not treat a tidy spreadsheet as proof the production inputs look the same.
Changing facts are a lookup job
SFT is a poor memory for text that is supposed to stay current. A wording PDF that renews, a returns policy that legal rewrote on Tuesday, a supplier feed that lands overnight — those are retrieval or process problems. Baking last week’s snapshot into weights teaches last week’s snapshot. People reach for SFT here because training feels more serious than a search index. Serious is the wrong test. Currency is the test.
If the workflow is “read this document and extract agreed fields”, the current document should be in the input. SFT then teaches how to extract, not what the latest PDF said. If the workflow is “answer from the company handbook”, and the handbook moves, retrieve the current page. You can still SFT the format of the answer later. Do not invert that order because someone booked GPUs.
Permission is not a hyperparameter
A cleared file list is part of the task. Training on uncleared correspondence, customer attachments, or supplier contracts you do not have the right to use is still a training run on uncleared files. Quantisation, adapters, and “we will only keep the adapter” do not create permission. The readiness review exists because those questions are the product.
Write who may see the train slice, who may see the unused pack, and where those files will sit after the run. If that paragraph cannot be written, stop. A later incident will not care that the loss looked healthy.
When SFT is the right conversation
Use SFT when the workflow is repeatable, pairs exist or can be made without a research programme, a named judge will defend the targets, unused test documents exist, and you will compare the adapted model with the current process. Insurance extraction and retail attributes are typical because a good answer can be written down. Open-ended customer advice is not typical. A chatbot that “knows the company” is usually retrieval plus a style guide, not a first SFT run.
SFT with LoRA is the common first training method once those conditions hold. That sentence is an order of operations, not a quality ranking. QLoRA enters only when memory is the constraint. Preference work enters only when two otherwise-correct drafts still fail a written bar. Distillation enters only after a teacher already meets that bar. None of those pages replace this one: if you cannot name the labels, do not book any of them.
Decision prompts, not scores.
When
Then
Pairs, a judge, permission and unused test documents exist
SFT can be the first training method. Write the schema into the brief.
Handlers disagree on letter tone and nobody owns the standard
Do not train. Write the tone rule. The run will freeze the fight.
Facts change weekly and the model is being asked to remember them
Retrieval or a process change first. SFT will freeze last week’s text.
The brief is only “fine-tune it” with a folder of files
Readiness review. A folder is not labelled pairs.
How people stretch the word fine-tune
A vendor chat product that stores a style on their stack is not this page. It can be a useful probe. It usually cannot travel with you as weights. Plan the schema so the same unused pack still works on the next base. See the closed-API comparison for the inference boundary.
A preference run that ranks two drafts is not this page. A full update that touches every weight is not this page. A public exam score is not this page. If the brief uses “fine-tune” for all four, rewrite the brief before anyone spends a specialist’s week labelling the wrong artefact.
Handover for an SFT run
The pack is the train-slice identifier, notes from unused test documents, the schema, the judge rule, the exact base checkpoint if you adapted one, the adapter or merged weights, and the licence line from the family page. Include the sentence that says what the model must refuse. Operators who only receive a weight file cannot tell a good extract from a confident guess.
If you cannot show the unused pack and the schema, you do not have an SFT handover. You have a training log. We will ask for the pairs before we treat the run as an engagement artefact.
SFT is the usual supervised form of fine-tuning. Preference methods such as DPO are a later or separate step. People say “fine-tune” for both, and for vendor UI styles. Write down which one you mean.
Do we need SFT if we already have RAG?
Not automatically. Retrieval helps changing knowledge. SFT helps stable behaviour and format. Many briefs need one, not both, as the first engagement. If the handbook changes weekly, do not start by baking last week’s PDF into weights.
Can we skip unused test documents if the examples look clean?
No. Clean training files are the easiest way to fool yourself. The reserved pack is how you still have a comparison with the current process.
How many pairs do we need?
Enough to represent the document families, languages and exception cases you will actually see — and still leave unused test documents untouched. We will not publish a magic count. A small clean schema beats a large disagreeing spreadsheet.
Can SFT add facts the documents never contained?
No. If the source quote is not in the input, the model is guessing. Teach restraint and an exception token, or retrieve the current source. Ranking two guesses later will not create a citation.
Does a lower training loss mean the extract is ready?
No. Loss on the train slice can fall while reserved documents still fail the schema. Compare with the current process on unused work, then stop or commission a fresh pack.
Handlers already agree on the facts but not the tone. Is that SFT?
Not yet. Tone without a written standard is two targets. Write the letter rule a conduct reviewer will sign. If they still disagree, you do not have a pair.
Does SFT make the model sovereign?
No. Labelled pairs do not move inference or change who can read logs. The sovereignty checklist is a separate operating document.
If you cannot name the labels, do not book SFT.
Bring the workflow and who judges a good answer. We will say whether supervised fine-tuning, retrieval or a readiness review is the first step.