XGBoost

Install mlbundle[xgboost,schema]. Models are stored as native in-memory UBJSON and loaded with load_model, never pickled.

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 = XGBClassifier(n_estimators=3, max_depth=2, n_jobs=1).fit(rows, [0, 1, 0, 1])
export(model, "xgboost.mlbundle", class_labels=[["cold", "warm"]])

bundle = load("xgboost.mlbundle")
assert isinstance(bundle.native_model, XGBClassifier)
assert bundle.predict_proba(rows) == model.predict_proba(rows).tolist()
assert bundle.native_model.predict(rows).tolist() == model.predict(rows).tolist()
print(bundle.class_labels)  # [['cold', 'warm']]

Class labels

XGBoost trains on encoded indices. Supply the original labels in probability-column order, nested even for one output. mlbundle binds them, with the prediction settings, to an attribute on a copy of the booster; your model object is not changed. Load verifies that the binding is consistent, not how the model was trained.

• Without class_labels=, encoded indices are used, or an existing binding.
• The native model’s predict keeps returning encoded indices; predicted_label in records maps the decision to your label.
• Regressors reject class_labels=.

Supported

• Exact XGBClassifier and XGBRegressor, one target.
• Native gbtree storage, including XGBoost 3.4.1’s dart option, which uses gbtree storage.
• Classifier objectives binary:logistic and multi:softprob.
• Regression objectives reg:squarederror, reg:absoluteerror, reg:pseudohubererror, reg:squaredlogerror, reg:logistic, count:poisson, reg:gamma and reg:tweedie.
• At most 100,000 trees and depth 14.

Structure is the complete list of JSON tree dumps with statistics, plus the native model record.

Rejected

Raw Booster objects, XGBRFClassifier/XGBRFRegressor, subclasses, ranking, survival and custom objectives, softmax-only objectives, gblinear, legacy dart storage and multiple targets.

What survives

Only configuration that native XGBoost I/O preserves is recorded and verified: the learner model parameters, objective, booster, the bound prediction settings (missing, importance_type, enable_categorical) and best-iteration attributes. Learning rate, max depth, seed and thread count do not survive native I/O and are neither recorded nor reconstructed.

The exact mapping is in Implemented XGBoost mapping.