Machine learning
Quorvyn Train
In developmentYour own AI model from your own data, guided and local
Quorvyn Train walks you in eight steps from a folder of pictures to a checked model: import, label, train, evaluate, export. Training and evaluation both run on your own machine.
Planned for: Windows 10 version 22H2 or newer, 64-bit · Linux
At a glance
- Status
- In development, not available yet
- Systems
- Windows and Linux. Windows 10 version 22H2 or newer, 64-bit.
- Processing
- On your device. Your content does not leave it.
- Prior knowledge
- Not needed. Eight steps lead to a finished model, and only expert mode shows the technical terms.
- Hardware
- Training runs on the CPU. Training on NVIDIA graphics cards is implemented, and the installation package for it will follow.
- Result
- An ONNX model, checked against the original, with a model card and sample code.
Features
- Eight steps instead of an empty screen
- Goal, data, labelling, data quality, training, evaluation, models, export. Every step says what it needs, what it produces and what comes next.
- Read a dataset as it already is
- Recognises folder-per-class, COCO, YOLO and Pascal VOC from the content itself rather than asking you for a format. A dataset of 60,000 images imports as well.
- A quality report before training
- Flags imbalanced classes, broken files, duplicates and images that are too small, before compute time goes into a dataset that already contains the problem.
- Labelling with undo
- Classes, boxes and polygons inside the application, with undo and redo. A correction is saved as a new version of the dataset and the old one stays.
- Training on your own machine
- The training service runs locally on 127.0.0.1, is started by the application and stops with it, never starts on its own with the system, and accepts only a randomly generated token per start. The images are handed to it as a file path rather than transferred.
- Evaluation in plain words
- Results in ordinary language rather than metrics, plus a gallery of the misses, meaning the images the model got wrong.
- An export that runs elsewhere
- An ONNX model, verified against the original, together with the class list, checksums, input format, a model card and example code in one package.
What Quorvyn Train deliberately does not do
These are not open tasks. They describe where Quorvyn Train stops, so that nobody builds on the assumption that it carries on.
- Graphics card later
- Training through CUDA on NVIDIA graphics cards is implemented, but the installation package does not include it yet. Until then Quorvyn Train trains on the CPU, with the runtimes that implies. It does not use Intel or AMD graphics.
- No resuming where a run stopped
- A cancelled run cannot be resumed where it stopped. You can, however, use a finished classification model as the starting point for a new run, as long as the classes stay the same.
- Images, not text and not tables
- Image classification, object detection and semantic segmentation are implemented. Text classification and tabular data are not.
- No fitness check
- Quorvyn Train verifies the export against the trained original, but that certifies no fitness for any particular use. Whether a model is good enough is decided by your task and not by our output.
- Desktop, not mobile
- Windows and Linux. There is no Android edition and no web interface, because training needs hardware a phone does not have.
Who Quorvyn Train is for
Quorvyn Train is meant for people who want to solve a concrete image problem without studying machine learning first. They might want to sort parts on a conveyor or spot recurring defects. They know what the result should look like, but not necessarily what a learning rate is.
That is exactly why learning rate, epochs, precision and recall do not appear in the interface. Anyone who wants to see them switches on expert mode. Anyone who does not never sees them and still trains a model.
Rule or model
Reading a dataset, the quality report and the export check are fixed rules, because file format, image size and checksum can be counted and compared. Quorvyn Train’s actual result, on the other hand, is itself a trained model that recognises what you teach it.
Under the Cyber Resilience Act, we have classified Quorvyn Train ourselves as a product with digital elements that counts neither as important nor as critical, so we assess its conformity ourselves, without a notified body. Quorvyn Train itself is not a high-risk system under the EU AI Act and is subject to no specific transparency obligations. What you use a model trained with it for is your decision. If you use it in an area the AI Act treats as high-risk, such as a safety component of a machine or the selection of job applicants, that use carries obligations of its own, and they fall to you.
Why this runs locally
Training data is often among the most valuable things a business owns, such as photographs from production or years of failure cases. The usual route asks you to hand exactly that over. You upload the data, have the model trained in the cloud and get it back.
Quorvyn Train does not ask that. The dataset stays where it is. The training service runs as a separate process on the same machine and receives the path to the images rather than the images. It starts when you start the application and ends with it.
Processed on your device
Quorvyn Train analyses your content locally. There is no transfer path through which it could leave the device.
Common questions about Quorvyn Train
When can I use Quorvyn Train?
Quorvyn Train is in development and not available yet. If you would like to hear from us once it launches, write to the address on Quorvyn Train's download page. We use your address for that one message and nothing else.
Do I need to understand machine learning?
No. Quorvyn Train leads in eight steps from a folder full of images to a trained model, and each step says what it needs, what it delivers and what comes next. Terms like learning rate, epochs or recall do not appear in the interface. Anyone who wants to see them switches on expert mode.
Which tasks can I solve with Quorvyn Train?
Image classification, object detection and semantic segmentation are implemented. A model from Quorvyn Train sorts images into classes, finds objects in them or marks out areas of an image pixel by pixel, for instance to sort parts on a conveyor or to spot recurring defects. Quorvyn Train does not handle text or tables.
Which systems and which hardware do I need?
Windows and Linux are planned. On Windows, Quorvyn Train will come through the Microsoft Store and needs Windows 10 version 22H2 or newer, 64-bit, and for Linux a direct download is planned. Training runs on the CPU for now, so it takes correspondingly longer. Training on NVIDIA graphics cards is implemented and will come with its own installation package. Quorvyn Train does not use Intel or AMD graphics. There is no Android version and no web interface, because training needs hardware that a phone does not have.
Do my images leave my computer?
No. The dataset stays where it is. The training service runs as a separate process on the same machine, starts with the application and ends with it. It receives the path to the images, not the images themselves.
Can I use the model outside Quorvyn Train?
Yes. Quorvyn Train exports an ONNX model that has been checked against the trained original, together with the class list, checksums, input format, model card and sample code. That check does not certify that the model is good enough for your task, though. Only the task itself can settle that.
Guides on this topic
Introduction
How can you train an AI model without coding?
What your own AI model learns, how many examples it needs, why training locally makes sense and where the limits are.
4 min read
Next step
Where and how you can obtain Quorvyn Train, which operating systems it is coming to and which route it ships through.
The button opens your mail program with the subject filled in. Without one, write to contact@quorvyn.de. We use your address for that one message and nothing else.