Training & Packaging

Create a dataset to train on, and train your model

Dataset Creation

Once you have data under all your labels and you think you are ready to go, we can create a dataset and train a model.

Open the Training section, and click the Training link. If you have more than one model on your project, make sure the model you want to work on is selected via the dropdown in the header of the page.

Click New Dataset to start a new dataset. Here you will need to set up your dataset:

Setting	Description
Name	This is inconsequential to function, mainly for differentiation.
Labels	Shows your taxonomy of labels. All are selected by default. You can disable labels for training.
Sources	Organize your uploads into sources (usually no need to change).
Include only accepted annotations	Excludes annotations marked as not ideal without deleting them.

You must run a check on the dataset query before submitting. This checks if you have the minimum required data to train. If you see any warnings or errors, expand them for necessary changes. When your check completes, you can create the dataset!

Datasets take a moment to build in the background, so you will be sent to the dataset overview page where it should be shown as "building." Refresh in a few moments, and it should be done.

Training

Once your dataset is built, you're ready to train! From the same "Training" view where you see your list of datasets, any dataset that is built and ready but not yet trained will have a "Train" button. Once clicked, training will begin, and you will need to wait for it to complete.

Estimated Train Time

Few-Shot takes anywhere from 8-20 minutes on average, but can grow with traffic and larger datasets.

Packaging

From the same Training view of your datasets, the Train button will become a Package button once the training is complete. You can click Details to get a summary of info on your model's training and the items included in your dataset as well. Once ready, hit Package, and you will be redirected to the package details and presented with options to download your model data.

If at any point you need help from the Passio team, please reach out to us at [e-mail]