Train a model

You need: the curator dataset role or higher.

Fine-tune on the annotations you have corrected, so the next pass through the loop starts from a better model.

Steps

1. Open the dataset's training page and start a training run.
2. Configure the run and submit it. Training is queued to the ai-worker Celery worker, so you can leave the page.
3. Watch the job card for progress.
4. When it finishes, the run, its metrics and the resulting model version are tracked in MLflow (port 5000 by default) and registered for reuse.

The registered model then becomes selectable in batch inference and in the model zoo.

Training modal → submit → job card progressing → resulting metrics / registered version.

Recording slot — model-zoo
Master-detail: model list → detail panel → sort by a performance stat → preview chips.
12–16 s · dark theme · 1280×800

Train on corrected annotations, not raw predictions

Because model output stays separable from human edits, you can tell the difference. Fine-tuning on the model's own unreviewed output teaches it its own mistakes.

Training without a GPU is impractical

It will run on the CPU, but not at a speed that supports iteration. See GPU and CUDA.