Composite pipelines

A fitted sklearn.pipeline.Pipeline whose final step is an exact XGBoost or CatBoost wrapper is exported as it is. The transformer prefix is stored with skops, the final estimator with its native adapter, and load rebuilds a real Pipeline after verifying both parts. Install mlbundle[sklearn,xgboost,schema] or mlbundle[sklearn,catboost,schema].

from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from xgboost import XGBClassifier

from mlbundle import export, load

rows = [[0.0, 1.0], [1.0, 0.0], [0.1, 0.9], [0.9, 0.1]]
model = make_pipeline(StandardScaler(), XGBClassifier(n_estimators=3, n_jobs=1))
model.fit(rows, [0, 1, 0, 1])
export(model, "pipeline.mlbundle", class_labels=[["cold", "warm"]])

bundle = load("pipeline.mlbundle")
assert bundle.predict_proba(rows) == model.predict_proba(rows).tolist()
assert bundle.native_model.predict(rows).tolist() == model.predict(rows).tolist()
print([member.value for member in bundle.members])
# ['manifest.json', 'model.skops', 'model.ubj', 'schema.json', 'structure.json']

How it is stored

Part	Member	Verified on load
Transformer prefix	model.skops	Step names, order, options and a fitted-state commitment
Final XGBoost estimator	model.ubj	The full XGBoost claims and tree structure
Final CatBoost estimator	model.cbm	The full CatBoost claims and tree structure

A single-step pipeline stores only the native leaf. The manifest reports the library as sklearn with serialization pipeline, and model.pipeline records the steps, the final estimator’s claims and the transformed-feature boundary between them. All-scikit-learn pipelines are not composite: they use one skops member and have no structure.

Labels, features and structure

• Labels follow the final estimator: XGBoost uses the labels you bind with class_labels=, and its native classes_/predict stay encoded; CatBoost keeps its native labels.
• Manifest features, schema, training dtypes and samples describe the original inputs.
• Trees address the transformed features, without attribution back to the original columns. mlbundle inspect prints the boundary’s width and names.
• Six members are enough for manifest, prefix, leaf, structure, schema and explicit samples.

Supported prefixes

Scalers, PCA, SimpleImputer (with indicators and empty columns), KBinsDiscretizer, PolynomialFeatures, univariate selectors, OneHotEncoder, OrdinalEncoder, identity or closed NumPy FunctionTransformer, nested native-free Pipeline and ColumnTransformer. Each needs a fitted, data-free proof of its output width.

Rejected

Transformers whose width cannot be proven, FeatureUnion, KNNImputer, slice, callable or scalar column selectors, pipeline caching (memory), metadata routing, and native estimators anywhere other than the final step.

Step names and outer options are reconstruction declarations; no name from the manifest ever selects loading code. The full representation is in Composite pipelines.