Fine-tuning · Lattice Séeb

A model adapted to your organization. Only for you.

We adapt a language model to your sector and vocabulary, as an applied research project with success criteria written before training.

Find out if your task justifies fine-tuning See how we specialize the model

Written success criteria
Evaluation report with method
Exclusive use inside Lattice

The problem

A general model knows a little of everything. Your operation needs more.

Fine-tuning is justified when the task is specific, repeats, matters to your operation and depends on your terminology, formats or judgment, and when there are examples to measure it. If looking up documents that change often is enough, document search is the better fit.

It answers confidently about what it does not know

A general model does not know your internal rules, your formats or your operation's nomenclature, and it still answers.

Your task repeats and the model fails the same way

Even with your documents at hand, it mixes up your terms, ignores your format or skips your judgment. That is where fine-tuning can do more than writing better instructions.

Nobody measures whether it worked

Without a baseline and an agreed evaluation set, there is no way to know whether the model improved your task.

Lattice Séeb

Specialized models.
Measured against your task.

Fine-tuning adapts a model to a specific domain: its terminology, formats, regulation and criteria. The goal is a model that is more accurate on your task, proven against a baseline, that runs inside Lattice: in our cloud, a private cloud or your server.

This line builds on Sintérgica AI's models: Lattice Na'at, with Mexican context, and Séeb, specialized by sector. Every project is quoted with a written research plan and success criterion.

Success criteria written before training
Baseline measured with general models before anything changes
Evaluation report with method
Exclusive use of the resulting model for your organization, inside Lattice

What information is needed

Operating manuals and internal procedures
Sector regulations (DOF, CNBV, CRE, COFEPRIS)
Technical terminology and glossaries
Template contracts, clauses and criteria
Protocols and practice guides
Historical reports, audits and rulings

Three things: process documents (like the ones on this list), examples solved by your team with the correct result, and an expert to review. The amount and quality are defined when the project starts: we inventory, curate and anonymize the data and build the evaluation set with your team.

Process

From your documents to an evaluated model

Each step has a deliverable, and metrics are set before we start.

Step 01

Definition and data

We agree on the task, the expected performance and how it will be measured. We inventory, curate and anonymize the data and build the evaluation set with you.

Deliverable: success criteria and evaluation set

Step 02

Baseline

Before training, we measure how general models solve the task, with and without search over your documents. We use the evaluation set: cases from your operation with their correct answer, which the model does not see during training.

Deliverable: baseline results

Step 03

Fine-tuning

We fine-tune the model on your cleaned documents, in controlled experiments.

Deliverable: model specialized in your domain

Step 04

Evaluation and deployment

We compare against the baseline with the same tests and the agreed metric: for example, per-field accuracy when extracting data, the right document when searching, or review by your experts with a rubric when drafting. If it meets the criterion, we deploy inside Lattice, in our cloud, a private cloud or your servers.

Deliverable: model in production and evaluation report

Comparison

General model vs. specialized model

What changes when the model is tuned to your task and measured with your data.

Aspect	General model	Specialized Lattice Séeb
Knows your internal terminology	What you give it in each instruction or through document search	Learned from your corpus
Result on your task	It is the baseline	Compared on the same tests; the report says whether it met the criterion
How it is measured	General tests	Your evaluation set, agreed before training
Where it runs	Provider's infrastructure	Our cloud, a private cloud or your server
Use of your data	Per the provider's terms	Only for your model, by contract
Who uses it	All of the provider's customers	Only your organization, inside Lattice
Mexican regulatory context	Depends on the model	Part of the training corpus

Industries

Where specializing a model pays off

Organizations with their own vocabulary, enough data and a concrete task.

Legal

Contract comparison, case file summaries, deadline tracking.

Financial

CNBV provisions, credit files, AML compliance support.

Energy

Regulatory and operating reports, technical quoting.

Healthcare

Administrative documentation within health regulation.

Government

Institutional-format letters, reports to other agencies.

Logistics and foreign trade

Customs regulation, TIGIE, file validation.

Deliverables

What you get at the end of the project

An evaluated, documented model running inside Lattice.

Model specialized in your domain, for your organization's exclusive use
Evaluation report against the baseline, with method
Evaluation set agreed with your team
Deployment in our cloud, a private cloud or your servers
Model and corpus documentation
Model update plan

Find out if your task justifies fine-tuning

Your data, only for your model

Your data is used only for this project and the resulting model is for your organization's exclusive use inside Lattice. Sensitive information is anonymized before training.

Exclusive use of the resulting model
Anonymization of sensitive information
Data terms in writing in the contract

FAQ

Questions about fine-tuning

No. Your data is used only for your project and the resulting model is for your organization's exclusive use inside Lattice. The terms are set in writing in the contract.

It's an applied research project: cost and duration depend on the volume and quality of your data. We give you both in writing in the proposal.

Document search retrieves information at query time and is useful when documents change often. Fine-tuning changes the model so it internalizes your vocabulary, formats and criteria. They are often combined; we decide with the evaluation.

Inside Lattice: in our cloud, a private cloud or your servers. With our models on your infrastructure or in the Mexico region of our cloud, information is processed in Mexico.

It depends on the task. We need examples to train on and others, kept apart, to evaluate. In the diagnosis we tell you whether your data is worth reviewing; when the project starts we inventory it and tell you whether it is enough and what is missing.

Next step

A model that knows your operation, measured with your numbers

In 45 minutes we tell you in writing what is worth specializing, what is not, and what it costs. The session has no cost.