scikit-learn

Install mlbundle[sklearn,schema]. Models are stored with skops.io.dumps and loaded with skops.io.loads against a trusted-types allowlist that lives inside mlbundle. The allowlist is never read from a bundle, and you cannot extend it.

from sklearn.ensemble import RandomForestRegressor

from mlbundle import export, load

rows = [[0.0, 1.0], [1.0, 0.0], [0.1, 0.9], [0.9, 0.1]]
model = RandomForestRegressor(n_estimators=3, random_state=0).fit(
    rows, [1.0, 2.0, 1.1, 1.9]
)
export(model, "forest.mlbundle", feature_names=["temperature", "humidity"])

bundle = load("forest.mlbundle")
assert bundle.predict(rows) == model.predict(rows).tolist()
structure = bundle.structure
print(structure["family"], structure["estimator_shape"])  # tree_ensemble [3]

Supported estimators

Fitted classifiers that expose predict_proba, and regressors whose feature and output counts can be recovered. Tested families include:

• decision trees, random forests, extra trees and gradient boosting;
• linear, regularized, robust and generalized linear models;
• SVMs, PLSRegression, naive Bayes, LDA and QDA;
• dense brute-force neighbors and NearestCentroid;
• MLPClassifier/MLPRegressor;
• MultiOutputRegressor/MultiOutputClassifier and RegressorChain.

Transformers that round-trip, including inside nested ColumnTransformer pipelines: SimpleImputer, KNNImputer, scalers, KBinsDiscretizer, categorical encoders, univariate selectors and FunctionTransformer with closed NumPy functions.

The exact class, attribute and function lists are in Prediction state; the tested configurations are in tests/test_sklearn_families.py.

Structure

Model	structure.json family
Single tree (tree_)	decision_tree
Forests, extra trees, gradient boosting	tree_ensemble
Models with coef_/intercept_	linear
Models with only feature_importances_	feature_importances
All-scikit-learn Pipeline, other families	null

scikit-learn structures also carry a prediction_state commitment and state_sha256, so the fitted state used for prediction is bound to the bundle. An all-scikit-learn pipeline has no structure, because version 1 has no mapping from its final step back to the input features, but its claims and state commitment still cover every step.

Rejected

Export fails with UnsupportedModelError, listing the untrusted types where skops finds them:

• custom estimators and subclasses;
• functions and types outside the closed lists, including lambdas and your own functions;
• compound fitted dtypes;
• classifiers without predict_proba;
• multilabel and ClassifierChain layouts;
• untrusted nested storage, such as fitted sparse neighbor arrays;
• HistGradientBoostingClassifier/Regressor, whose TreePredictor storage is not on the allowlist.

You cannot opt in to trusting a rejected type. Rewrite the model with supported components, or keep it out of the bundle.

Versions

The allowlist is audited and tested against scikit-learn 1.9.1 with skops 0.16.0, not every version the >=1.6 range allows. Unknown fitted state fails. Use the same scikit-learn version for export and load; a failed load under a different declared version raises EnvironmentMismatchError to say so.