CatBoost

Install mlbundle[catboost,schema] (CatBoost 1.2.10 or newer; it brings matplotlib and plotly). Models are stored as native CBM and loaded in memory.

from catboost import CatBoostClassifier

from mlbundle import export, load

rows = [[0.0, 1.0], [1.0, 0.0], [0.1, 0.9], [0.9, 0.1]]
model = CatBoostClassifier(
    iterations=3, depth=2, verbose=False, allow_writing_files=False
)
model.fit(rows, ["cold", "warm"] * 2)
export(model, "catboost.mlbundle")

bundle = load("catboost.mlbundle")
print(bundle.class_labels)  # [['cold', 'warm']]
assert bundle.predict_proba(rows) == model.predict_proba(rows).tolist()

CatBoost owns its labels: classes_ is verified on load and class_labels= is rejected, even when equal.

Supported

Exact fitted CatBoostClassifier and CatBoostRegressor, one target, numeric features.
SymmetricTree (oblivious trees) only.
Classifier losses Logloss, CrossEntropy and MultiClass.
Regressor losses RMSE, MAE, Quantile, Huber, Poisson and MAPE, including their native parameter strings.

Complete trees, feature borders, NaN handling, scale and bias, and feature importances are carried in structure.json and verified on load. predicted_label uses CatBoost’s own class prediction, including its threshold.

Rejected

class_labels=; categorical, text and embedding features; Depthwise and Lossguide growth; MultiClassOneVsAll and other losses; multi-target and uncertainty models; rankers, bare CatBoost, subclasses; custom Python objectives or metrics; unknown wrapper attributes.

Categorical features round-trip natively in CBM, but their CTR storage has an unchecked native loader, so mlbundle rejects them before export and again before native load. Encode categories yourself, or use a scikit-learn prefix in a composite pipeline.

The wrapper recovery, native-parser limits and in-memory serializer are specified in CatBoost mapping.