Build and inspect datasets

Create specialist or multimodal datasets around the outcome the system must understand.

Source reviewed 2026-09-23. Availability depends on your installation, permissions, and compatible runtimes. A supported path is not a guarantee of model quality or production readiness.

Outcome

Create a named collection of media and labels. Review and validation are still required before it can be used for training or testing.

Start here

What you need to know first

Check the inputs below before starting. If you are new, begin with the first-project guide. A dataset holds media and labels; a model produces results; a deployment makes a selected model version callable. Creating one does not create the others.

Bring these inputs

Images, video frames, or connected object storage
An objective-specific annotation schema

Expected output and limits

Immutable dataset versions
Split manifests, statistics, and lineage

Core workflow

01
Choose the objective
Select detection, segmentation, classification, phrase grounding, VQA, or captioning.
02
Bring the evidence
Upload images or clips, connect storage, or select existing data. To generate images, provide a scene description, objects, and variations; a dataset name is not a generation brief. Review the server quote and accept its maximum before starting. Managed generation costs $0.20 per image saved; unused reserved credit is released. Connected providers bill you directly.
03
Inspect dataset health
Review splits, classes, balance, annotation coverage, dimensions, and spatial statistics.
04
Version the contract
Create the next immutable version before downstream training or evaluation.

What this surface supports

Specialist and VLM schemas
Aggregated dataset statistics
Train/validation/test splits
Working versions, frozen inputs, and lineage

Expert section

Contracts, signals, and failure modes

Use this section when you are defining acceptance criteria, automating the surface, or reviewing whether its output is safe to promote downstream.

Quality and operating signals

Coverage and class balance
Duplicate and corrupt assets
Spatial and dimensional distributions

Common failure modes

Train/test leakage
Schema drift between versions
Representative classes with too few examples

Expert release checklist

Inputs and dependencies are pinned to immutable versions.
Acceptance metrics include critical classes and operating slices.
Failure, retry, cost, and rollback behavior are understood.
The resulting artifact has an owner and a downstream review path.

Engineering safeguards

Schema follows the selected objective
Warnings expose imbalance and coverage gaps
Downstream runs bind exact versions