Skip to content
⁨NDDev AI⁩, home page
Community
Language: English
Display
Theme
Motion

Training AI models on your own data

NDDev trains neural networks on company data: we fine-tune language models with LoRA and QLoRA for your style, format and terminology, and train classification and computer vision models on your examples. You receive the model weights, training code and a report with metrics on a held-out set. It suits companies with labelled data. We have worked since 2019.

Who it is for

Companies that have built up examples: requests with correct categories, documents with filled-in fields, photos with marked defects, correspondence in the house style.

  • LoRA and QLoRA for language models
  • Models on PyTorch and Transformers
  • Experiments tracked in MLflow
  • Weights and training code are yours

When a model should learn from your data

First we check whether training is needed at all. If the task is answering from documents, a search over your knowledge base is usually enough. Fine-tuning pays off when the model needs a strict format, narrow terminology or your house style and prompts give unstable results. For classification, field extraction and image tasks, a small dedicated model often runs faster and cheaper than a large language model.

  • a data audit and labelling plan
  • labelling and cleaning of examples
  • a baseline model for comparison
  • LoRA or QLoRA fine-tuning, or training a dedicated model
  • an experiment log in MLflow
  • a report with metrics on a held-out set
  • model weights, training code and run instructions

Most of the work goes into the data. We clean and label examples, split them into training, validation and held-out sets and record the result of a simple baseline model. We fine-tune language models with LoRA and QLoRA on agreed hardware and train other models with PyTorch, Transformers or YOLO. Every run, with its parameters and metrics, is logged in MLflow.

If it is unclear whether the target quality is reachable, we start with a research stage: we try methods from simple to complex and share interim results straight away. The outcome is a working model or a reasoned finding that the target cannot be reached with this data.

Training stages

  1. Task and metric

    We agree what the model does and which metric will tell us it works.

  2. Data

    We check the volume and quality of examples and label or extend them where needed.

  3. Baseline

    We train a simple model or try prompts so there is something to compare against.

  4. Experiments

    We select the architecture and parameters and log every run in MLflow.

  5. Validation and handover

    We measure the model on the held-out set, deploy it on your side and hand over the weights and code.

Questions about training on your own data

How much data do we need?

That becomes clear after the data audit. We look at how many examples you have, how consistent they are and whether they are labelled, and tell you plainly whether they are enough or labelling comes first.

Can the model be trained on our servers?

Yes. We run training on your hardware or in an agreed cloud, and the data and resulting weights stay there.

Do you guarantee the model’s accuracy?

We are accountable for the agreed process, experiments, report and documentation. We fix the target metric before we start and measure it on a held-out set the model never saw in training. If the target cannot be reached with this data, you learn that at the research stage.

Who owns the trained model?

Rights to the weights and training code are transferred by agreement and set out in the contract.

How much does training cost?

The starting price in the price block covers our work. The total is shaped by the volume of data and labelling, the size of the model and the number of experiments. Computing power and external labelling are billed separately at cost.

from $8,900

Free warranty
80 days

Prices are starting prices for information only and are not a public offer. The final price is fixed in the specification.

Language models and company knowledge

Describe the task and the data you have

Tell us what the model should do and how many examples you have collected. We will assess whether the data is enough and suggest a plan.

We reply within 24 hours

Assistant

Loading the assistant…