Models trained on the way
your industry actually writes.

A general model has never seen your part codes, your abbreviations, your clause numbering or the shorthand your team types into a ticket at two in the morning. When that costs you accuracy, the answer is to adapt a model to your material. When it does not, the answer is cheaper, and we will say so first.

This is for you if

Start here

Five things to try, cheapest first

Training a model is the last step, not the first. Most accuracy problems are solved somewhere higher up this list, for a fraction of the cost, and we work down it in order rather than starting at the bottom because it sounds more impressive.

1

Write the instruction properlyTry first

Be specific about the output format, give it the edge cases, show it four or five worked examples. Unglamorous, and it fixes a large share of what people bring to us as a training problem.

2

Give it your documents at the moment it answersUsually the answer

Retrieval, rather than training. The model looks your facts up instead of memorising them, so when a policy changes you update a document rather than retrain anything. For knowledge, this beats training almost every time.

3

Use a smaller model for the repetitive part

Classifying, extracting and routing rarely need the largest model. Moving that work to a small one usually cuts cost and latency at once, with no loss on the task.

4

Adapt a model to your materialWhere we train

Now it is worth it. Teaching a small open model your formats, your vocabulary and your house style, from examples your team has already produced. This is what most people mean by a custom LLM.

5

Teach it which answer is betterRarely needed

Preference training, where the model learns from pairs your experts ranked. Reserved for tone and judgement, where there is no single correct answer to point at.

Nobody trains a model from nothing any more, and you should be careful of anyone offering to. Building a base model costs millions. Everything worth doing starts from an open model that already exists.

Where this is used

Six jobs where adapting a model earns its cost

Documents

Reading forms nobody else has seen

Industry paperwork with layouts and codes a general reader gets wrong. A model taught on a few thousand of your own examples reads them reliably. Typical content: shipping documents, lab reports, claim forms.

Language

Vocabulary that means something different here

Where a word means one thing in ordinary use and another in your trade, and getting it wrong changes the decision. Typical content: clinical, legal, insurance and engineering terms.

Cost

One high volume task, done cheaply

A small trained model that does one job as well as a large general one, at a fraction of the cost per item and several times the speed. Typical work: classifying, routing, tagging, extracting.

In your product

A feature that has to sound like you

Drafting, replying or summarising in your own house voice, trained on what your team has already written and approved. Typical content: past replies, published copy, style guides.

Language

Languages the big models handle poorly

Regional languages, mixed-script writing and technical Arabic are noticeably weaker in general models. Adaptation closes much of that gap. Typical content: local-language tickets, forms and correspondence.

Search

Teaching search what counts as similar

Training the retrieval side rather than the answering side, so that two of your part numbers are recognised as related. Cheap, quick, and often the biggest single gain in a retrieval system. Typical content: your catalogue, your codes.

Send us fifty examples of the task and we will tell you whether training would help.

Services

Everything the work involves, in order

01

Agree what "better" means, in a number

Before anything else, a set of your real cases with the correct answer agreed and written down. Without it nobody can tell whether training helped, and every later argument becomes a matter of opinion.

02

Measure the model before any training

Several models, properly instructed, scored against that set. Often the honest report is that a better instruction closed most of the gap and training is not worth it. That is a real outcome and we deliver it when it is true.

03

Build the training set out of work you already did

Your approved decisions, corrected extractions and past replies are training data. Finding it, cleaning it and getting the rights and privacy straight is the largest part of this job, and the part most people underestimate.

04

Fill the gaps in that data

Real examples cluster around the easy cases. We generate and check additional examples for the awkward ones, because a model trained only on easy cases fails exactly where it matters.

05

Keep a set of examples the model never sees

Held back from the start and used only for the final score. Without this, a model that has quietly memorised the answers looks perfect right up to the day it goes live.

06

Choose the base model and the training method

Which open model to start from, and which technique to adapt it with. We use the light methods that train a small add-on rather than the whole model, because they are far cheaper and can be undone.

07

Train it, and record how

Every setting, every version of the data, every result, logged. If a model six months from now behaves oddly, you can see exactly what produced it and rebuild it.

08

Check it did not get worse at everything else

Training on one task can quietly damage others. We test general ability alongside the target task, because a model that reads your invoices perfectly and has forgotten how to write a sentence is not a win.

09

Put it into production behind the same door

Served through the same internal endpoint as everything else, so the application does not know or care which model answered. That is what lets you compare the trained model against the general one on live traffic.

10

Retrain when it is worth it, not on a schedule

We watch accuracy on live work and retrain when it drops or when your material changes. Retraining on a calendar burns money and occasionally makes things worse.

Our stack

The tools, standards and methods we use

Listed so a technical reader can see the approach is ordinary and reproducible, and ask us why we picked one over another.

Models we start fromOpen weight models with licences that allow commercial use. We check the licence before recommending one
LlamaMistralQwenGemmaPhi
How we adapt themLight methods that train a small add-on rather than the whole model. Far cheaper, quicker, and reversible
LoRAQLoRAPEFTSupervised fine-tuningFull fine-tuning where justified
Teaching preferenceFor tone and judgement, where there is no single right answer to point the model at
DPOPreference pairsReward modelling
Making a small model as good as a big oneThe large model produces worked examples, the small one learns from them and then does the job
DistillationTeacher and student modelsSynthetic data generation
Building and labelling the dataThe largest part of the work. Labelled once, checked by more than one person where it is ambiguous
Label StudioArgillaAgreement between reviewersDeduplicationPersonal data removal
Running the trainingStandard libraries on rented or your own graphics cards. Nothing here is proprietary to us
Hugging Face TransformersTRLAxolotlUnslothDeepSpeedPyTorch
Keeping track of what produced whatEvery run recorded with its data version and settings, so any model can be rebuilt or explained
Weights and BiasesMLflowModel registryDataset versioningModel cards
Scoring the resultOn your own task set, plus general tests to confirm nothing else broke
Held-out test setPrecision and recall per fieldlm-evaluation-harnessBlind human comparison

Example

Where the accuracy actually came from

Example

A logistics company was reading customs paperwork with a large general model and getting about 71% of fields right. They asked us to train a custom model, and had budgeted for it.

We built the scoring set first, then worked down the ladder. Rewriting the instruction to state the document type, define each field and show four filled examples took accuracy to 84%, in two days and at no running cost. Adding a lookup of their own commodity codes at the moment of answering took it to 91%, because most remaining errors were codes the model had no way of knowing.

Only the last stretch justified training. A small open model, adapted on about four thousand of their own previously corrected documents, reached 96% and ran at roughly a fifth of the cost per document of the large general model they started with.

The point is the order. Had we trained first, the same 96% would have arrived later, cost more, and nobody could have said which part of it was the training. Two thirds of the gain came from steps one and two.

An example, with invented figures, to show the method. Your numbers will differ.

FAQ

Questions we get asked before starting

How much data do we need?

Fewer examples than people expect, but they have to be good. A few hundred well chosen ones often beat tens of thousands of noisy ones. What matters most is covering the awkward cases, not the volume.

Can we train it on our confidential data?

Yes, and training can run inside your own environment so nothing leaves. The resulting model holds patterns from that data, so who may use it is a real question and we settle it in writing before starting.

Will it stop making things up?

It reduces certain errors and does not remove the behaviour. If the goal is factual accuracy, retrieval with citations is the right tool and training is not. We would rather tell you that than sell you the bigger job.

Who owns the trained model?

You do. The trained weights, the training data and the code to rebuild it. Subject to the base model's own licence, which we check and tell you about before we start rather than after.

How long does it take?

Usually six to ten weeks, and most of that is building and checking the data. The training itself is often a single overnight run.

What if it does not beat the general model?

Then the scoring set says so plainly, and you have a measurement worth having rather than a belief. It happens, most often when the real problem was retrieval or the instruction all along.

Next step

Show us where the model keeps getting it wrong.

Fifty examples of the task, including the ones it fails on. We will tell you on the call whether this is a training problem, a retrieval problem, or an instruction that needs ten minutes of work.

Ahmedabad, India. We work with teams in the US, UK, Europe, Singapore and the Gulf, and we are used to the time difference.