LLM fine-tuning for work your team already does.

Who this page is for

Read this page when someone in the building has already said “we should fine-tune” and you need a commercial fork, not a method. Typical readers are a sponsor holding a training budget, a specialist who would have to label the work, and a buyer who was shown a demo and asked to book GPUs. The page is not a tutorial for adapters and it is not a ranking of foundation families.

If the live question is whether a smaller model is even a candidate, open the SLM article. If the live question is where the running system will live, open inference. If the files cannot yet be used, open readiness. Stay here when the spend decision is train, retrieve, change the desk, or stop.

The fork this page is selling

Buyers arrive here after someone said “we should fine-tune.” The useful reply is not a training date. It is a fork: pay to adapt behaviour on a stable job, pay to retrieve changing text, or fix the workflow so a model would have something honest to learn. Those three spend the same quarter differently. Mixing them in one statement of work is how pilots stall and how a later slide pretends three unfinished jobs were one method.

Work your team already does is the filter. If handlers extract named fields from a policy family every day, that job can be described, sampled and compared. If the request is “make the model understand our company,” there is no job yet. We will not invent one so a training run has a title. A title without a desk is how a marketplace sells a week.

The commercial question is whether an adapted model beats the current path enough to justify training, review and the later operating cost. That question only exists after you write the current path down. A demo that looks fluent is not that writing. Fluency is cheap. A baseline you can defend is not.

When adaptation is the engagement

Book a fine-tuning conversation when three things are already true. The task is the same shape tomorrow as it was last month. Someone in the business can mark a right extract, a right class, or a right letter without inventing policy on the spot. You can spare a pack of real cases that the optimiser will never see, so the comparison is not a rehearsal of the training file.

Typical work in that shape: pulling agreed fields from a stable wording family, classifying a submission into queues you already run, filling catalogue attributes from a schema that does not change weekly, or drafting correspondence that a named reviewer still signs. The model is being asked to behave like the desk, not to replace the authority behind the desk. Authority is an insurance-page problem. This page only needs you to notice when the brief has left the desk and entered a verdict.

If those three are missing, the first paid week should not be a training week. It should be a readiness review, a retrieval design, or a workshop that writes the task. Training on disagreement copies the disagreement. Training on last quarter’s facts teaches last quarter’s facts. Both are engagements we will refuse to dress up as adaptation.

• One named job, not “our documents” as a category.
• A current baseline you can describe without a vendor slide.
• A judge who already owns the output in production.
• Permission to use the examples for training and for the unused comparison pack.
• A written point at which a human must take the case.

When retrieval or a process change should go first

Retrieval is the conversation when the text that must be right changes faster than you would want to retrain. Product prices, live stock, a new underwriting circular, a returns policy that marketing rewrote on Tuesday: those belong next to the prompt, not inside the weights. Adapting a model so it “remembers the handbook” is how you ship a confident stale answer. Confidence is not currency. Currency is a source that is still true this week.

A process change is the conversation when the desk itself is the problem. If two specialists mark the same clause differently, you do not have labels. If nobody can say what a good letter contains, you do not have a target. If the queue that would consume the model’s output does not exist, you are buying a file that nobody will open. Fix the desk, then decide whether the desk is worth teaching to a model.

Some briefs need both later. The first engagement is still one of the three. We will write which one, and what would make us switch, before anyone transfers a corpus. A statement of work that says “fine-tune and RAG and redesign the queue” is three invoices pretending to be one agile phase.

Insurance: extract the wording you already handle

A binder team already pulls named fields from a wording family and pastes them into a workbench. The fields are agreed. The exception path is “send to a senior.” Volume is high enough that a fluent general model still wastes time on format and restraint. That is an adaptation candidate: teach the extract shape and the refusal to invent a missing limit. The engagement question is whether that waste is expensive enough, and stable enough, to justify a training week rather than a better template.

It is the wrong first booking if the real pain is last month’s circular that changed a definition. That circular belongs in a retriever or a process that opens the current file. Training last year’s PDFs will not see the new definition. It is also the wrong booking if the brief says the model should decide coverage. That is a claims engine. We will not take that as a fine-tuning scope. See the insurance page for the review boundary. This page only records the fork: stop, or split the extract from the verdict.

Retail: attributes on a schema you already own

A catalogue team already fills material, fit, and care from supplier text into a schema merchandising owns. The schema is stable for a season. Reviewers already reject bad fills. That is an adaptation candidate: teach the attribute list and the refusal to guess a fibre that is not in the copy. The engagement question is whether the current fill path — humans, rules, or a prompted general model — is the baseline you will compare against.

It is the wrong first booking if the pain is this week’s promo price or live size availability. Those values live in merchandising systems. A model that “learned last season’s prices” is a liability. Retrieval or a hard lookup is the first design. Adaptation, if you still want it, is for the stable attribute language — not for the number that changed this morning.

What a useful brief contains before we talk training

Name the job in one sentence a specialist would recognise. Describe what happens today, including the ugly exceptions. Say how you will judge a better answer without asking us for a score we have not measured. Say where the model must run, because that constraint can kill a family before a single epoch. Say who reviews silence — the cases the model must not complete. Those five sentences are the brief. A folder of PDFs is not a brief.

That is enough to choose among adaptation, retrieval, a smaller-model comparison, or a readiness review. It is not enough to start training. Training also needs permissions, a three-way split, and a written go / no-go. Those live on the data-readiness page. Do not paste a confidential corpus into the contact form to skip that step. Skipping it is how a training invoice becomes the first artefact by accident.

Model artefacts, licensing, deployment and any later support are written into the engagement before work begins. Insurance and retail teams usually buy the readiness week first. That is not delay for its own sake. It is how we avoid charging you for a run that cannot be evaluated. A brief that asks us to skip readiness because “we already have lots of files” is still a brief that has not met the five sentences.

How we scope without opening a training cluster

The first call is a scoping call. We ask for the job, the baseline, the judge, the environment, and the evidence a buyer needs before anyone trains. We do not ask you to upload the book. We do not promise a method from a homepage tag. LoRA, full updates, preference work and distillation are technique pages. They are not this page’s product. Booking a method before the fork is how a team buys QLoRA when they needed a circular in a retriever.

If the fork points at adaptation, the next paid artefact is usually a readiness review: permissions, coverage, sensitive-data bounds, and a split that keeps the comparison pack unused. If the fork points at retrieval or a process change, we will say so in writing and stop. That refusal is part of the practice. A training invoice is not a consolation prize for a brief that was not an adaptation job.

Success, when we do train, is a comparison you can defend: the adapted model versus the baseline on cases the run did not see, plus latency and operating cost if those change the business case. We agree the criteria before the run. We do not retrofit a win from a chat that looked good in a notebook. A notebook is a tool. It is not an engagement outcome.

What lives on other pages on purpose

Supervised fine-tuning, adapters, preference methods and compression are technique briefs. They answer how a run is shaped once the engagement is adaptation. Open them after this page says “yes, train.” Do not start there because a slide said QLoRA. The QLoRA page is a memory method, not a quality tier, and it is not this commercial fork.

Family pages answer licence and serving questions for a named base. If counsel said “open source” or someone booked Llama by habit, read those pages and the openness comparison. This page will not paste Community Licence clauses so that a chatbot can quote them twice. Quoting them twice is how commercial URLs stop owning one query.

Inference and sovereignty pages answer where prompts go and what you can take with you. Those questions often arrive wearing a fine-tuning costume. Take the costume off. If the real ask is “the prompts must not leave our tenancy,” you are not yet in a training conversation. You are in an operating conversation that might later allow a training conversation.

What we will not train

We will not train a model to decide coverage, settlement, or a customer’s legal position. We will not train on examples you cannot permit. We will not treat a public invoice demo as evidence for your book. We will not sell “sovereign AI” as a side effect of a LoRA file. Those refusals are how this page stays the engagement fork instead of becoming a method shop.

We will not invent a task so the quarter has a pilot. If the only brief is “use our documents,” the engagement is to write the job, not to start an epoch. That sentence disappoints some sponsors. It is also the sentence that keeps this site citable. A citable stop is still a commercial outcome.

Questions we hear first