Models

Beta

A model groups everything you train for one task: its trained versions, its training history, and its settings. You train models on dataset versions of your reviewed images.

Model access

Model creation and training are in private beta and available only to organizations with access, on any plan. Without access, the app shows Private beta and a Request model access link, which opens the model access request page.

Model types

Model type	Availability
Object detection	Private beta
Instance segmentation	Coming soon
Image classification	Coming soon
Language model (LLM)	Coming soon

Object detection learns to find and locate objects in images with bounding boxes. The other types are shown in the app but cannot be selected yet.

Find your models

Open Models in Annota Studio to see the models you can access. Use Show to filter by All, Mine, Shared with me, or Organization. Each row shows the model task and its latest training.

Create a model

1. In Models, select Create model.
2. Enter a Model name, for example “Parts detector”, and an optional Description.
3. In Model type, choose Object detection.
4. Select Create model.

You choose the dataset and base model later, when you train a model.

The model page

Click a model to open it. The page has three sections:

• Versions: the trained versions of the model.
• Training history: every training run, with its status and metrics.
• Settings: name, description, and deletion.

Use Share in the page header to give members, groups, or the organization access.

Model versions

Each successful training run creates one new, immutable model version. The Versions table shows each version number, the run that produced it, its status, and its Archive size. Open a version to see the training run that created it.

To download a version, select the download icon on its row. The archive contains:

• The trained weights, as the best and the last checkpoint of the run.
• A manifest and the split of images between training and validation.
• A README that records the configuration, metrics, and class mapping.

Annota does not run trained versions inside the app. Download the archive to use the weights in your own environment.

Note

Models are fine-tuned from third-party pretrained checkpoints, so the downloaded weights are subject to the AGPL-3.0 license unless an enterprise license from the checkpoint provider applies. The training dialog links to the full license.

Edit or delete a model

1. Open the model and select Settings.
2. Change the name or description and select Save changes.
3. To delete the model, select Delete model in Danger zone and confirm.

Deleting a model removes its versions, training history, and saved weights. The dataset versions it was trained on stay available and are released, so you can delete them. Wait for an active training run to finish, or cancel it, before you delete a model.

Sharing and permissions

Models are private to their creator until shared.

Access	What you can do
Can view	See versions and training runs, follow progress, and download versions
Can edit	Also train new versions, cancel or duplicate runs, edit the name and description, and share with members and groups
Owner	Also delete the model and set organization access

Organization admins can also delete models. See Sharing and access.

Next steps

• Train a model
• Dataset versions
• Review assets