refit_full

Description

Specifies whether to refit the best AutoML model (fitted on train data) on the full (train+val+test) data.

After we complete all the requested experiments by training on the train split and using the validation split for early stopping, Kumo will take the best-performing model and retrain it using all available data, which includes the training, validation, and holdout test data splits. This allows us the final model used to generate predictions to use as much and as recent data as possible while avoiding leakage in AutoML.

NOTE: This option can be used together with refit_trainval option. If refit_trainval is also set to true, Kumo reports the holdout results based on a model re-trained on train+val data. The Batch Predictions will still be based on a model refitted on the full data.

Supported Task Types

• All

Default Values