Output Notebook

Export contract

The workflow exports exactly one file: a self-contained Jupyter notebook.

The notebook is intentionally simpler than the package runtime. It contains only the code required for the agreed workflow path and does not import agentic_automl when executed.

Notebook structure

The exported notebook is organized around five explicit workflow functions:

  1. preprocessing(dataset_path, target, features, transformation_kwargs)

  2. split_data(transformed_data)

  3. train_model(train_split)

  4. hyperparameter_optimization(train_split)

  5. validate(test_split, trained_model)

Each section includes only the code relevant to the selected workflow path. The exporter prunes unused helpers so the notebook reflects the actual agreed workflow rather than a generic package runtime.

What the notebook includes

  • relevant input information

  • selected policies

  • metric summaries

  • model and baseline comparisons

  • hyperparameter competition output when tuning was run

  • plots relevant to the chosen task

  • a rerunnable minimal runtime embedded directly in the notebook

Path handling

The notebook resolves the dataset path relative to the notebook location, so the export stays portable within the project tree.

Validation story

Validation compares the trained model against the strongest simple no-feature baseline:

  • classification uses a class-prior baseline from the training target distribution

  • regression uses a target-only constant baseline from the training target distribution

Final validation can also include the tuned-versus-untuned comparison and the hyperparameter competition dashboard.