Curate & manage: datasets, slices, and lineage

A useful training set has a purpose: improving a particular scenario, balancing coverage, or testing a failure hypothesis. Curation makes the selection explicit so you can compare the next model with the previous one.

Understand the working pieces

• Datasets organize source data such as images, video, point clouds, and MCAP recordings.
• Projects define labeling work and its configuration.
• Tasks are the units of work whose status and results you inspect.
• Selections and slices identify the examples relevant to your use case, using the metadata and query capabilities available for that data.

A selection like “night scenes with pedestrians” is useful only if the necessary attributes exist and are reliable. Decide how those attributes will be recorded and checked.

Select for a reason

Start with the failure or coverage gap you want to address. Then identify candidate sequences, remove unsuitable inputs, and check that the required labels have passed review. Separate training and evaluation data deliberately to avoid measuring performance on examples the model has already seen.

The right amount of data depends on the task. Compare the benefit of a targeted selection with the cost of collecting and labeling it; there is no fixed percentage that works for every project.

Preserve the handoff

Record the source dataset and project identifiers, the selection criteria, the quality gate, and the export used for training. Keep that record with the model and evaluation results. Review history and source references help explain a label; an explicit release record helps reproduce a training run.

Inspect projects with the SDK

This read-only example lists the first page of projects visible to your API key. It does not create a slice or change project status:

from avala import Client

with Client() as client:
    for project in client.projects.list(limit=10):
        print(project.uid, project.name, project.status)

Use the SDK reference to inspect the project and its tasks before creating a training-data export.

Next: Train: from reviewed data to model →