Custom LoRA model
A model that speaks like your business: your jargon, your formats and your tasks. We build the dataset, train the model and measure whether it really improves.
LoRA fine-tuning adapts an open model to your business (vocabulary, format, tasks) and leaves it running on your hardware. For organisations with highly specific knowledge and good-quality examples. It takes 4 to 8 weeks, with your own dataset, training and before-and-after evaluation, and the resulting model is yours.
Generic models know a little about everything, but they fail exactly where your business is different: your sector’s terminology, the exact format of your reports, the method you use to explain things, the tone you take with your clients. It can be corrected with long instructions and a private RAG search, and often that is enough. When it isn’t, the solution is to teach the model with examples: fine-tuning.
Doing it in the cloud means uploading your best examples to a third party, and those are usually your most valuable knowledge and often contain personal data subject to the GDPR or information under professional secrecy. With LoRA, training can be done on your own hardware: the dataset and the resulting model stay in-house, and the model then runs on your server without paying per token.
If what you want is for the model to “know” your data, such as your catalogue, your clients or regulations that change, fine-tuning isn’t the right tool: it forgets them, mixes them up and goes out of date. That’s what RAG is for. Nor do we recommend it if there are no good-quality examples and nobody to validate them, or if a commercial cloud model already does the task well and your data isn’t sensitive. For most businesses we see in Barcelona, fine-tuning is the second step, not the first.
For Nuzic, an educational music and mathematics project from Barcelona, we built the dataset factory (examples with reasoning, trick questions that contrast with classical music theory and multi-turn conversations, in three different tones), a curation website where the team validates each example, and the LoRA fine-tuning of a custom model. That’s why we know the real work is in the data, and that’s where we put the effort.
An assistant that knows your case files and templates, runs inside the firm and never exposes a single piece of client data.
Less time typing and more time with the patient: transcription and summaries that stay inside the clinic.
Your technical know-how, searchable in seconds and without leaving the company network.
Tutors and assistants that teach with your method and your syllabus, without sending student data to the cloud.
A field agent that speaks Catalan, knows your farm and keeps the logbook for you.
Less than you might think: with LoRA, training fits on an office GPU or a DGX Spark, and the fine-tuned model is then served on your server like any other. The real work is in preparing a good dataset. We work from Barcelona, remotely or at your office.
RAG gives the model up-to-date information with every question; fine-tuning teaches it how to answer: style, format, vocabulary and typical reasoning. For data that changes, RAG; for behaviour, fine-tuning. Often the best option is to combine them.
Usually, from a few hundred good examples for a specific task. If you don’t have them, we generate them from your documents and your team validates them; quality matters more than quantity.
You do. The dataset, the LoRA adapter and the merged model stay on your server. Before we start, we check the base model’s licence with you to make sure it allows the commercial use you need.
LoRA fine-tuning adapts an open model to your business (vocabulary, format, tasks) and leaves it running on your hardware. For organisations with highly specific knowledge and good-quality examples. It takes 4 to 8 weeks, with your own dataset, training and before-and-after evaluation, and the resulting model is yours.