Skip to navigation
Data LabOnboarding

Onboarding

Understand how a datasource becomes a dataset for training.
View as Markdown

Onboarding prepares an ingested datasource for a prediction task. It applies the target, selected features, prediction type, preprocessing, sharding and encoding choices that make the data ready for training.

The output is a dataset. The datasource describes the imported data; the dataset represents that data after the onboarding choices have been applied.

How onboarding works

1

Datasource

Onboarding starts from an available datasource created by ingestion.

2

Configuration

The onboarding configuration defines the prediction task and how to prepare the data, including missing-value handling and categorical preprocessing.

3

Job

The onboarding job applies the configuration to the selected datasource using processing compute.

4

Dataset

A successful job creates or updates a dataset. Inspect its fields, distributions and lineage before selecting it for training.

Onboarding resources and their relationships

An onboarding job combines a datasource, onboarding configuration and processing compute to produce a dataset.

Initial and incremental onboarding

An Initial job creates a dataset. An Incremental Update job updates an existing dataset from a datasource containing additional data. The configuration must match the job type; an optional Increment label identifies the update.

Updating a dataset does not change a model already trained from it. Run a new training job when you need a model built from the updated data.

Find onboarding resources in the app

Under Data Lab > Datasets, use onboarding configurations for the preparation choices, onboarding jobs for execution, and datasets for the outputs. A configuration is reusable; each job records a particular run and its resource references.

Next steps

Create an onboarding configuration, run an onboarding job, then inspect the dataset.