Executed examples

These standalone scripts run during each documentation build. Every page shows the executed source, its output and plots derived from the returned fitted state. All data are synthetic. The short fits demonstrate the API; they do not establish posterior convergence, calibration or a recommended component count.

All model inputs use samples × features. Numeric outputs use float64 arrays unless an output container is selected explicitly. Training scores_ contains joint variational means; transform(new_rows) computes conditional projections under the fixed fitted state. Missingness and scoring targets are described with each example.

Example	What it demonstrates
Missing dense observations	Reconstruction, independent new rows and full latent covariance
Sparse observations	Stored observed zeros, probes and an equivalent dense observation pattern
Mixed-data Pipeline	Raw masks, schema, row CV/GridSearchCV and decoding new rows in input units
Cost versus probe RMS	Common held-out cells, separate metric units and explicit capacities
Original-variable cell validation	Training-only preprocessing, categorical scores and numeric units
Plotting utilities	Scree, component loadings and explained variance from the fitted model

These examples use the 0.5 API. From its matching source checkout, install python -m pip install '.[plot]' to run the scripts. Published 0.4 packages use the earlier estimator interface. Optional table output also needs the data extra for pandas or the polars extra for Polars. The complete Pipeline tutorial executes both supported table containers and describes metadata routing and scoring.

The migration guide explains the explicit feature-first legacy adapter and compatibility choices. The source-collected examples in the estimator, options and preprocessing references are executed in CI too.