Examples¶
These examples are small enough to run locally and use real package interfaces. They illustrate composition and validation rather than benchmark performance or model accuracy.
| Workflow | What it demonstrates |
|---|---|
| Raw input to prediction | Fit preprocessing and a model together; predict without supplying derived columns. |
| Sibling features and ordered dependencies | Build features in successive specifications and drop temporary columns. |
| Parallel transformed and unchanged features | Combine a scaler and FeaturePassthrough in a ColumnTransformer. |
| Lazy feature preparation | Defer feature evaluation and collect explicitly at the model boundary. |
| Custom Python feature functions | Define map_elements callbacks, compose stages, and pickle/restore fitted preparation. |
| Schema drift | Reject missing columns and dtype changes, normalize explicitly allowed changes, and guard prediction in a pipeline. |
| Cross-validation | Clone and fit the full pipeline per fold. |
| Persistence and warning filters | Save a fitted pipeline and choose whether mismatches warn, fail, or are silenced. |
Prediction and evaluation gallery¶
| Question | Workflow | Data |
|---|---|---|
| Can order value predict a fee? | Held-out split, schema validation, feature selection, scaling, regression | Generated locally, 80 orders |
| How stable is a full pipeline across folds? | Clone and fit the entire preparation path per fold | Local order fixture |
| Where should a lazy query become eager? | Deferred Polars preparation and estimator boundary | Local order fixture |
Copy each complete Python block into a file and run it in the installed environment. Assertions and marked displayed output are validated by the documentation checker.