RELEASE NOTES

Version 0.6.0

Breaking changes

• Drop support for Python 3.9, which reached end-of-life in October 2025. The minimum supported version is now Python 3.10
• NGBoost.score now returns the negative of the average scoring rule (the mean log-likelihood for LogScore), so that higher is better as scikit-learn expects. Previously, GridSearchCV and cross_val_score without an explicit scoring argument selected the worst model (issue [#418], PR [#423])
• Base learners are now seeded from the model random_state, so fits are reproducible and independent of n_jobs. Output for the default base learner shifts relative to 0.5.x (PR [#417])

New features

• Add optional n_jobs parameter to fit base learners in parallel with threads (PR [#416])
• Support a distinct base learner for each distribution parameter, including scikit-learn nested parameters and feature importances for all-tree configurations (issue [#338], PR [#411])
• Add validation_fraction and early_stopping_rounds to NGBClassifier and NGBSurvival for API parity with NGBRegressor (issue [#402], PR [#407])
• Report fitted status to scikit-learn's check_is_fitted via __sklearn_is_fitted__, fixing NotFittedError in Pipeline.predict on sklearn >= 1.8 (PR [#429])
• Set scikit-learn classifier metadata on NGBClassifier so that scikit-learn correctly identifies it as a classifier (PR [#413])
• Handle array-like sample weights positionally and validate them at fit time; weights with wrong length, shape, NaN, or inf now raise ValueError (PR [#425])
• Document saving and loading models with skops and test the round trip (PR [#427])

Bug fixes

• Fix the CRPScore metric of Normal, NormalFixedMean, and LogNormal, which was off by a factor of the predicted scale. Fits using CRPScore with these distributions now produce correct natural gradients (issue [#421], PR [#424])
• Fix the scale gradient of the Laplace CRPScore, which dropped a 0.75 * scale term (PR [#419])
• Fix censored scores and gradients for Exponential and LogNormal, which swapped censored and observed rows and had sign and extra-term errors (PR [#420])
• Fix the log-scale Fisher information of TFixedDf and Cauchy, which was off by a factor of 4 * var (PR [#422])
• Reset validation iteration metadata on refit so that re-calling fit on the same model behaves identically to fitting a fresh one (PR [#426])
• Keep random_state as given and seed a fresh generator in fit, so that get_params returns the original value and clones behave identically (PR [#417])
• Replace the deprecated friedman_mse criterion with squared_error for the default tree learner (issue [#408], PR [#409])

Documentation and CI

• Build the user guide with Jupyter Book 2 in GitHub Actions and deploy on push to master (PR [#428])
• Remove the old prebuilt docs site generated by jupyter-book 0.x in 2020 (PR [#430])
• Run user guide notebooks during the docs build with --execute, replacing removed load_boston calls with load_diabetes (PR [#431])
• Mark observed rows as events in the survival examples (PR [#436])
• Run the test matrix once per commit and cancel stale PR runs via concurrency groups (PR [#444])
• Fix scikit-learn wheel resolution on Python 3.14 (PR [#414])