Read and reuse AutoML reports

AutoML saves a run-level README.md and individual model reports in its results directory. The leaderboard compares validation scores. Keep final held-out test results separate from the scores used to select models.

For details behind individual predictions, see model explainability. Once you have evaluated a model, reload it for predictions or create a prediction web app.

Open the HTML report

After running the customer-churn tutorial, load the saved run and create the HTML report:

from supervised import AutoML

model = AutoML(results_path="AutoML_CustomerChurn/automl")
model.report()

Open AutoML_CustomerChurn/automl/README.html in a browser. In a notebook, leave model.report() as the last expression in a cell to display it. Click a model name to inspect its settings, validation metrics, and available explanations.

The concrete-strength walkthrough and churn walkthrough show actual reports from recorded beta runs. For classification, check the threshold next to each metric: scores shown at different thresholds need not occur together.

Use a structured report

report_structured() provides text or data for analysis tools and LLM workflows. Using the loaded model above:

markdown_report = model.report_structured()
payload = model.report_structured(format="dict")
payload_json = model.report_structured(format="json")

Each call creates or updates report_structured.json in the results directory. The default output is Markdown text. For one model, pass its exact saved name:

detailed = model.report_structured(model_name="3_Linear")
print(detailed)

3_Linear exists in the recorded churn tutorial; use a name from your own leaderboard for another run. A selected-model report includes detailed metrics, available feature-importance summaries, and fairness details when present.

Inspect explanations

Available plots depend on the algorithm and explain_level. See explainability for learning curves, permutation importance, SHAP, and model blueprints. Importance describes model reliance on an input; it does not establish that changing the input causes an outcome to change.

Keep the saved model artifacts

Reports help explain the result, but they are not a substitute for the model files. Preserve the complete results directory for prediction and reload, including preprocessing and ensemble dependencies. See saving and loading.