Changelog

Unreleased

Release highlights

Changed

Added

Removed

Fixed

0.17.0 (2026-04-30)

Changed

• Breaking change: EstimatorReport.metrics.add() now supports scorers, in the scikit-learn sense, i.e. callables with the signature (estimator, X, y_true, **kw) -> score. Callables with the signature (y_true, y_pred, **kw) -> score are no longer supported directly: users should convert those to scorers using sklearn.metrics.make_scorer(). See #2814 by @auguste-probabl.

Added

• Project now accepts strings for the workspace argument. See #2832 by @auguste-probabl.

0.16.0 (2026-04-20)

Release highlights

• Breaking change: Computing custom metrics on reports is now done by calling EstimatorReport.metrics.add() followed by summarize(). custom_metric() has been removed, and summarize() no longer accepts callables or keyword arguments; only names of registered metrics are supported (which includes the default metrics such as accuracy for classifiers). Example usage is available here. See #2736 by @auguste-probabl.

• Reports now expose a report.diagnose() method that automatically detects common modeling issues such as overfitting and underfitting.

Changed

• Breaking change: cache_predictions no longer accepts n_jobs or response_methods on estimator, cross-validation, and comparison reports. It now infers and caches the relevant prediction outputs automatically. See #2677 by @cakedev0.

• Breaking change: Computing custom metrics on reports is now done by calling EstimatorReport.metrics.add() followed by summarize(). custom_metric() has been removed, and summarize() no longer accepts callables or keyword arguments; only names of registered metrics are supported (which includes the default metrics such as accuracy for classifiers). Example usage is available here. See #2736 by @auguste-probabl.

• Python 3.14.4 is now supported, while Python 3.10 support is now deprecated. See #2771 by @thomass-dev.

Added

• summarize() (and equivalent methods on CrossValidationReport and ComparisonReport) now accept scikit-learn metric names without the neg_ prefix. For example, "mean_squared_error" can be passed instead of "neg_mean_squared_error" and is resolved automatically. See #2735 by @direkkakkar319-ops.

• Reports now expose a report.diagnose() method that automatically detects common modeling issues such as overfitting and underfitting.

• EstimatorReport and CrossValidationReport now expose get_state() / from_state() helpers to support more robust serialization and deserialization across skore versions than plain pickling. See #2741 by @cakedev0.

• mae(), map() (and equivalent methods on CrossValidationReport and ComparisonReport) are now available as default regression metrics, exposing mean absolute error (MAE) and mean absolute percentage error (MAPE). See #2765 by @direkkakkar319-ops.

0.15.0 (2026-04-02)

Changed

• Breaking change: Display.plot() (and all concrete displays) now returns a matplotlib.figure.Figure instead of storing plotting artifacts on the display. The attributes figure_, ax_, and facet_ are no longer set. Use fig = display.plot(...) and then fig.axes, fig.show(), or rely on the figure’s rich representation in notebooks. #2660 by @glemaitre.

• Breaking change: EstimatorReport now requires X_test. If X_test is omitted (None), construction raises ValueError. #2673 by @glemaitre.

• Breaking change: binary classification metrics and curve displays no longer require pos_label to be set, and skore no longer infers it implicitly. When pos_label is left unset, precision and recall metrics, as well as ROC and precision-recall curves, now expose both classes instead of failing. #2663 by @cakedev0.

Added

• Add rich HTML representation for the different reports. See #2632, #2651, #2658, and #2659 by @glemaitre, @jeromedockes and @GaelVaroquaux.

0.14.0 (2026-03-19)

Release highlights

• skore.evaluate() and skore.compare() are new top-level dispatcher functions that create the appropriate report class from an estimator in a single function call. See #2573 by @raotalha71 and #2604 by @glemaitre.

• skore can now integrate with MLflow by passing mode="mlflow" and the new tracking_uri option to Project. Example usage is available here. See #2527 by @cakedev0.

Changed

• Breaking: The filtering and display arguments previously accepted by .summarize(), e.g. aggregate, flat_index, favorability, have been moved to frame(), for every report. See #2536, #2545, and #2566 by @auguste-probabl.

• Breaking: Display.set_style() now returns None instead of self. Method chaining such as display.set_style(...).plot() is no longer supported. See #2579 by @direkkakkar319-ops.

• Breaking: pos_label can no longer be overridden when calling a metric or a display; it must be set when creating the report. See #2588 by @jeromedockes.

• Breaking: The name parameter of ComparisonReport.create_estimator_report() has been renamed to report_key. See #2561 by @GaetandeCast.

Added

• PermutationImportanceDisplay now supports a level parameter to select which level of a multi-index DataFrame to use for feature names. See #2565 by @GaetandeCast.

• CoefficientsDisplay now supports aggregation via an aggregate parameter in frame(), following the same pattern as ImpurityDecreaseDisplay. See #2552 by @MuditAtrey.

• PermutationImportanceDisplay and ImpurityDecreaseDisplay now support parameters select_k and sorting_order, with the same behaviour already available in CoefficientsDisplay. Passing select_k=0 now raises a clearer error. See #2591 by @GaetandeCast.

Removed

• Breaking: The data_source="X_y" option has been removed from report classes; "train" and "test" are now the only options. See #2537 by @jeromedockes.

Fixed

• CrossValidationReport.metrics.summarize now raises a NotImplementedError upon data_source="both". See #2548 by @auguste-probabl.

• Passing sparse matrices to data accessors now raises a NotImplementedError instead of crashing with an unhelpful error. See #2543 by @KaranSinghDev.

0.13.1 (2026-03-05)

Release highlights

• ComparisonReport now supports permutation importance through permutation_importance(). See #2511 by @GaetandeCast.

• ImpurityDecreaseDisplay.frame() now supports aggregation through an aggregate parameter. See #2539 by @MuditAtrey.

Fixed

• The figure-level legend is no longer cut off when saving the figure produced by PredictionErrorDisplay.plot(). See #2530 by @MuditAtrey.

0.13.0 (2026-02-26)

Release highlights

• Help menus accessed through help() methods now render as interactive HTML in Jupyter notebooks and IPython environments. See #2316 by @glemaitre.

Changed

• Breaking: Reports no longer accept clustering models and only support supervised learning tasks (classification and regression). See #2489 by @GaetandeCast.

• Breaking: The mode parameter is now required when creating a Project in “hub” mode. See #2401 by @thomass-dev.

• DataFrame column names in reports now consistently use singular form (e.g., “model” instead of “models”). See #2392 by @Sharkyii.

• Exceptions raised during the CrossValidationReport fitting process are no longer caught by skore. See #2462 by @glemaitre.

Added

• The documentation now includes a short example illustrating local usage of skore.Project. See #2481 by @glemaitre.

• The documentation now includes a short example explaining the design philosophy of the skore API. See #2480 by @glemaitre.

• The report URL is now printed after uploading to Skore Hub. See #2488 by @rouk1.

• Reports now support setting pos_label after initialization, enabling users to change the positive class for binary classification metrics. See #2438 by @glemaitre.

• CrossValidationReport now supports permutation feature importance through permutation_importance(). See #2370 by @glemaitre.

• ComparisonReport now supports mean decrease in impurity (MDI) feature importance through impurity_decrease(). See #2387 by @auguste-probabl.

• ConfusionMatrixDisplay now supports threshold_value="all" to display confusion matrices for all available thresholds. See #2463 by @glemaitre.

Fixed

• Tab completion now works correctly for reports in IPython. See #2427 by @jeromedockes.

• PredictionErrorDisplay now supports multioutput regression models. See #2434 by @GaetandeCast.

• CoefficientsDisplay now uses separate y-axes for each model when comparing models with different features. See #2403 by @GaetandeCast.

• Reports no longer crash when y is provided as a list or tuple. See #2433 by @cakedev0.

• skore.train_test_split() is now more robust to large inputs. See #2404 by @cakedev0.

• Data normalization in the data accessor is now more robust when converting various input formats to DataFrames. See #2440 by @cakedev0.

• Project names are now limited to 64 characters when interacting with Skore Hub. See #2454 by @auguste-probabl.

• An error is now raised earlier and more clearly when a report is pushed to Skore Hub without a pos_label. See #2436 by @glemaitre.

• Empty project names are no longer allowed when interacting with Skore Hub. See #2453 by @auguste-probabl.

• pandas.Timestamp objects no longer crash uploading a report to Skore Hub. See #2484 by @thomass-dev.