How can you train an AI model without coding?

What an AI model learns

An AI model learns from examples how to sort data or estimate values. What that means in detail depends on the kind of data. An image model answers, for instance, whether a part is fine or damaged, finds objects in an image or marks out areas pixel by pixel. A text model sorts enquiries or reports into a category. A model for tables estimates from measurements or order data which case applies or which value to expect. In every case the model does not learn from rules someone writes down but from examples where the right answer is already known. It is therefore only as good as those examples, and that is where the real work begins.

How many examples it takes, and which

There is no fixed number. The more alike the cases the model has to tell apart, the more examples it takes. For simple distinctions in images a few dozen to a few hundred images per class are often enough. For text and tables it depends heavily on how clearly the classes show in the data. What usually matters more than the amount, though, is how the examples are put together.

If one class has ten times as many examples as another, the model mostly learns the common one. If every picture of a class was taken on the same table in the same light, it may learn the background instead of the object. The same happens with tables when a column already gives the answer away, such as a status that was only entered after the outcome was known. The model then looks excellent in testing and fails in use. Duplicate entries, broken files and wrong labels make the result worse without it being obvious straight away. And some of the data does not belong in training at all. It is held back to test the finished model, because only that shows whether it also copes with cases it has never seen.

Why training locally makes sense

The usual route goes through the cloud. You upload the data, train there and get the model back. Training data, though, is often exactly what you would rather not hand over, such as pictures from production, customer enquiries or measurement series collected over years. When you train locally, it stays on your own computer. The price is computing time. Image and text models in particular take considerably longer on an ordinary computer without a suitable graphics card than in a data centre, while models for tables often train quickly on the CPU as well.

What works without coding and what does not

Today you can train a usable model without programming skills, as long as an application guides you through the steps, from importing and labelling the data through checking the dataset and training to evaluation and export. One decision no application takes off your hands, though, which is whether the model is good enough for your task. That shows in the cases it gets wrong and in what a mistake costs in real use. If a model is used in an area the EU AI Act classifies as high risk, such as selecting job applicants, that use also comes with obligations of its own.

Where Quorvyn Train helps

This path from your own data to a checked model is what we are building Quorvyn Train for. The application guides you through it in eight steps and does without technical vocabulary in the interface. Training and evaluation run on your own computer, and the training service is given only the path to the images, not the images themselves.

An honest description includes the limits. Quorvyn Train is in development and not available yet, and it is planned for Windows and Linux. So far only tasks with images are implemented, namely image classification, object detection and semantic segmentation. Quorvyn Train does not handle text or tables yet. So far it trains on the CPU. Training on NVIDIA graphics cards is implemented, and the installation package for it will follow. A cancelled run cannot be resumed where it stopped. And an exported model is not fit for a particular use just because Quorvyn Train produced it.