Train a model

Beta

A training run teaches a model from a ready dataset version. When the run succeeds, it adds a new version to the model that you can download.

Training is in private beta: only object detection on image datasets is available, and only to organizations with model access.

Before you start

You need:

• Model access for your organization, and edit access to the model.
• Edit access to the image dataset you train on, because the run keeps that dataset version.
• A ready version of that dataset with at least two approved images, where every label is used on an approved image.

Only approved images are used. Approved images without annotations teach the model what is not an object. Polygons become the boxes that enclose them. If your version does not meet the requirements, review more images and create a new version.

Start a training run

1. Open the model and select Train new version. You can also select Train new version on a ready version in the dataset’s Versions page, then choose a model or create one.
2. In Image dataset, choose the dataset. In Dataset version, choose a ready version. The dialog shows how many approved images and classes it holds.
3. In Compute, keep CPU. GPU is coming soon.
4. In Base model, choose the pretrained starting point. Larger bases need more memory and training time. Nano is recommended for quick CPU runs.
5. Optionally, open Advanced settings to change the configuration.
6. Select Train new version.

The run is queued and continues in the background. Select View training to follow it. A model can have only one active run at a time.

Training settings

Setting	Values	Default	What it does
Epochs	1 to 100	20	Maximum number of passes over the training images
Image size	320 px or 640 px	320 px	Resolution images are resized to for training
Batch size	1 to 16	4	Images processed together in one step
Validation images (%)	10 to 40	20	Share of images held back to evaluate the model
Random seed	0 to 4,294,967,295	42	Makes the split and training more repeatable

The validation share is a target. Identical images always stay on the same side of the split, and every label stays in training. The same seed improves repeatability but does not guarantee identical results.

Follow a run

Open Training history on the model page and click a run to see its details: status, current phase, completed epochs, duration, dataset version, base model, and configuration.

Status	Meaning
Queued	The run is waiting to start.
Running	The run is in progress.
Canceling	The run is stopping after a cancel request.
Completed	Training finished and a new model version was created.
Failed	Training stopped with an error. No version was created.
Canceled	The run was canceled. No version was created.

While a run is active, its phase shows what it is doing: Waiting for compute resources, Preparing the dataset, Training the model, Evaluating the model, or Saving model weights and results.

The person who started the run gets an email when it succeeds or fails.

Metrics

Latest reported metrics shows the latest values the run reported, and charts show their trend once there are at least two measurements.

Metric	Meaning
mAP50-95	Overall detection quality across overlap thresholds
Precision	How many detections are correct
Recall	How many labeled objects are found
Training loss	How well the model fits the training images

Higher is better for mAP50-95, precision, and recall. Lower loss usually means a better fit. Compare runs that use the same data and validation settings. A missing metric is not zero.

Cancel or repeat a run

• To stop an active run, select Cancel training, then Stop training. No model version is created.
• To start a new run with the same settings, open a run and select Duplicate configuration.

When a run fails

The run page explains the failure. Common causes:

• The dataset version contains invalid annotations or missing image files.
• There are not enough reviewed images: review more annotated images and create a new dataset version.
• The run exceeded its time limit: try fewer epochs or a smaller image size.

Credits and data

• Training currently uses no credits during the private beta.
• The dataset version used by a run cannot be deleted while the model exists.
• Downloaded weights are subject to the license described in Models.

Next steps

• Models: download trained versions.
• Dataset versions
• Review assets