Onboarding configuration | Umnai

An onboarding configuration defines how a datasource is prepared for training. An onboarding job runs that configuration and produces a dataset. Create the configuration, run the job and check the resulting dataset, then create a workspace to organise model development.

Before choosing onboarding settings, review Prepare data, especially the guidance on missing values, categorical values and targets. Use those preparation choices to guide preprocessing.

Before you begin

Have an AVAILABLE datasource, its data schema and a processing compute configuration ready. Identify the prediction type and target fields for the model before creating the onboarding configuration. Use one target for single-output or multi-class predictions, and more than one target for multi-valued output predictions.

The Datasets area has three tabs: Configurations stores onboarding settings, Onboarding Jobs lists processing runs, and Datasets lists the resulting datasets.

Configure initial dataset onboarding

Open Data Lab > Datasets > Configurations, then select Create.
Enter a Configuration name and select Initial for the first onboarding of a dataset.
Choose the Prediction type for your task. Use Classification for a class or category, or Regression for a numeric value. The example uses Classification.
Select the Data schema, then select the model’s target field or fields in Targets to match the intended outputs. The payments fraud example uses fraud_flag.
Review the encoding, sharding, masking and additional settings before selecting Create.
Return to Configurations and verify that the saved configuration is present before creating a job.

Initial onboarding configuration with fraud_flag selected

The example shows One hot encoding max categories as 64 and Infrequency threshold as 1. These control categorical encoding and the treatment of infrequent categories. The Shards control shows 256m, a memory limit per shard. Use its mode selector to choose a memory limit or a shard count, and review Sharding strategy for your prediction task.

Review processing settings

Review the masking fractions and additional preprocessing controls below. Keep the defaults unless your data requires a specific change. Decide how to handle unseen categories, missing values and categorical whitespace before saving the configuration.

Control	Example value or purpose
Mask category fraction	0.37
Mask fraction	0.8
Mask median fraction	0.15
Mask rows fraction	0.3
Handle unseen categories	Enabled in the example; enables masking for categories not seen during training.
Impute	Disabled in the example; controls whether missing values are imputed.
Encoding strategy	Select Stream or Shard as the encoding strategy for the onboarding configuration.
Strip categories	Disabled in the example; removes leading and trailing whitespace from categorical values.

Masking fractions and additional processing settings

Choose categorical and missing-value handling

For encoding examples and the relationship between schema settings and onboarding, see Encoding.

Control	How to choose
One hot encoding max categories	Threshold beyond which categorical fields switch from one-hot to target encoding. The UI starts at 64; review the number of distinct categories in your data.
Infrequency threshold	Groups infrequent categories under an infrequent-category representation. The UI starts at 1.
Handle unseen categories	Enables masking for categories absent during training. Keep prediction-time input conventions consistent with this choice.
Strip categories	Removes leading and trailing whitespace. Enable it when that whitespace is incidental, after checking whether trimming would merge meaningful categories.
Impute	Controls imputation of missing values. Decide whether missingness should be cleaned before ingestion or handled during onboarding.
Encoding strategy	Choose Stream or Shard when the workflow requires an explicit strategy; otherwise leave the optional choice unset.

Choose sharding

Sharding divides the data for processing and training. Shards has two modes: Max. Memory Per Shard takes a memory value such as 256m or 1g; Shards takes a positive count. Switching mode resets the value, so check it again before creating the configuration.

Sharding strategy	Intended use
Proportionate	Retain class proportions when dividing classification data. This is the default for classification when no strategy is supplied.
Disproportionate Minority Oversampling	Increase representation of minority classes for an imbalanced classification task.
Disproportionate Majority Undersampling	Reduce representation of majority classes for an imbalanced classification task.
Random	Divide data using random assignment.
Random Approximate	Use approximate random sharding. This is the default for regression when no strategy is supplied.

Choose a disproportionate strategy deliberately and evaluate against a population that represents your intended use. Changing the balance used in training does not change the population against which performance should be assessed.

Understand masking settings

Masking settings determine how observations and categorical values are sampled for masking during preprocessing. Their values are fractions, not percentages entered as whole numbers.

Control	Purpose
Mask category fraction	Fraction used to select categorical columns for masking.
Mask fraction	Fraction used to mask the selected categorical column.
Mask median fraction	Fraction sampled per categorical value.
Mask rows fraction	Fraction sampled from each shard and target for masking.

Start with the form’s values unless the preprocessing task calls for a deliberate change. Review the resulting dataset’s distribution and special categories before training.

Create an incremental configuration

Choose Incremental Update when preparing settings for an existing dataset. The form retains masking, missing-value and additional preprocessing controls while the dataset’s existing context supplies the prediction task. Initial-only settings, including the target, prediction type and sharding setup, are not entered again.

Save this configuration separately and select it in an incremental onboarding job. Use Initial when you need a new dataset with a different prediction task.

Review and reuse a configuration

Open the saved configuration and verify its type and preprocessing choices. For an initial configuration, check the target and prediction type. Select a configuration matching the job type when running onboarding; saving it alone does not process a datasource.

Next steps

Continue to Onboarding jobs.