The skore API

This example illustrates the consistent API shared by skore reports and displays. Reports expose the same accessors (data, metrics, inspection), and each method that produces a visualization returns a Display object. All displays implement a common interface: plot(), frame(), set_style(), and help().

Reports share the same accessor structure

EstimatorReport, CrossValidationReport, and ComparisonReport all expose the same accessors where applicable:

• data: dataset analysis

• metrics: performance metrics and related displays (e.g. ROC, confusion matrix)

• inspection: model inspection (e.g. coefficients, feature importance)

The data accessor is only available on EstimatorReport and CrossValidationReport because when comparing models, the input data can be different and thus one can access the underlying reports to inspect the data.

Calling a method on these accessors returns a Display object. The same pattern holds across report types, so once you know one, you know them all.

Minimal setup: one report and one display

We build a simple EstimatorReport and use it to show how accessors return displays and how those displays behave.

from sklearn.datasets import load_breast_cancer
from sklearn.linear_model import LogisticRegression
from skore import EstimatorReport, train_test_split
from skrub import tabular_pipeline

X, y = load_breast_cancer(return_X_y=True, as_frame=True)
split_data = train_test_split(X=X, y=y, random_state=42, as_dict=True)
estimator = tabular_pipeline(LogisticRegression())

report = EstimatorReport(estimator, **split_data)

╭────────────────────── HighClassImbalanceTooFewExamplesWarning ───────────────────────╮
│ It seems that you have a classification problem with at least one class with fewer   │
│ than 100 examples in the test set. In this case, using train_test_split may not be a │
│ good idea because of high variability in the scores obtained on the test set. We     │
│ suggest three options to tackle this challenge: you can increase test_size, collect  │
│ more data, or use skore's CrossValidationReport with the `splitter` parameter of     │
│ your choice.                                                                         │
╰──────────────────────────────────────────────────────────────────────────────────────╯
╭───────────────────────────────── ShuffleTrueWarning ─────────────────────────────────╮
│ We detected that the `shuffle` parameter is set to `True` either explicitly or from  │
│ its default value. In case of time-ordered events (even if they are independent),    │
│ this will result in inflated model performance evaluation because natural drift will │
│ not be taken into account. We recommend setting the shuffle parameter to `False` in  │
│ order to ensure the evaluation process is really representative of your production   │
│ release process.                                                                     │
╰──────────────────────────────────────────────────────────────────────────────────────╯

Data accessor: report.data.analyze() returns a display

The data accessor provides dataset summaries. Its analyze() method returns a TableReportDisplay.

data_display = report.data.analyze()
data_display.help()

Every display implements the same API. You can:

• Plot it (with optional backend and style):

data_display.plot(kind="dist", x="mean radius", y="mean texture")

You can set the style of the plot via set_style() and then call plot():

data_display.set_style(scatterplot_kwargs={"color": "orange", "alpha": 1.0})
data_display.plot(kind="dist", x="mean radius", y="mean texture")

• Export the underlying data as a DataFrame:

data_display.frame()

	mean radius	mean texture	mean perimeter	mean area	mean smoothness	mean compactness	mean concavity	mean concave points	mean symmetry	mean fractal dimension	radius error	texture error	perimeter error	area error	smoothness error	compactness error	concavity error	concave points error	symmetry error	fractal dimension error	worst radius	worst texture	worst perimeter	worst area	worst smoothness	worst compactness	worst concavity	worst concave points	worst symmetry	worst fractal dimension	target
0	12.89	13.12	81.89	515.9	0.06955	0.03729	0.02260	0.01171	0.1337	0.05581	0.1532	0.4690	1.115	12.68	0.004731	0.013450	0.016520	0.005905	0.01619	0.002081	13.62	15.54	87.40	577.0	0.09616	0.1147	0.11860	0.05366	0.2309	0.06915	1
1	13.40	20.52	88.64	556.7	0.11060	0.14690	0.14450	0.08172	0.2116	0.07325	0.3906	0.9306	3.093	33.67	0.005414	0.022650	0.034520	0.013340	0.01705	0.004005	16.41	29.66	113.30	844.4	0.15740	0.3856	0.51060	0.20510	0.3585	0.11090	0
2	12.96	18.29	84.18	525.2	0.07351	0.07899	0.04057	0.01883	0.1874	0.05899	0.2357	1.2990	2.397	20.21	0.003629	0.037130	0.034520	0.010650	0.02632	0.003705	14.13	24.61	96.31	621.9	0.09329	0.2318	0.16040	0.06608	0.3207	0.07247	1
3	17.75	28.03	117.30	981.6	0.09997	0.13140	0.16980	0.08293	0.1713	0.05916	0.3897	1.0770	2.873	43.95	0.004714	0.020150	0.036970	0.011100	0.01237	0.002556	21.53	38.54	145.40	1437.0	0.14010	0.3762	0.63990	0.19700	0.2972	0.09075	0
4	20.58	22.14	134.70	1290.0	0.09090	0.13480	0.16400	0.09561	0.1765	0.05024	0.8601	1.4800	7.029	111.70	0.008124	0.036110	0.054890	0.027650	0.03176	0.002365	23.24	27.84	158.30	1656.0	0.11780	0.2920	0.38610	0.19200	0.2909	0.05865	0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
564	14.64	15.24	95.77	651.9	0.11320	0.13390	0.09966	0.07064	0.2116	0.06346	0.5115	0.7372	3.814	42.76	0.005508	0.044120	0.044360	0.016230	0.02427	0.004841	16.34	18.24	109.40	803.6	0.12770	0.3089	0.26040	0.13970	0.3151	0.08473	1
565	14.45	20.22	94.49	642.7	0.09872	0.12060	0.11800	0.05980	0.1950	0.06466	0.2092	0.6509	1.446	19.42	0.004044	0.015970	0.020000	0.007303	0.01522	0.001976	18.33	30.12	117.90	1044.0	0.15520	0.4056	0.49670	0.18380	0.4753	0.10130	0
566	11.04	16.83	70.92	373.2	0.10770	0.07804	0.03046	0.02480	0.1714	0.06340	0.1967	1.3870	1.342	13.54	0.005158	0.009355	0.010560	0.007483	0.01718	0.002198	12.41	26.44	79.93	471.4	0.13690	0.1482	0.10670	0.07431	0.2998	0.07881	1
567	19.81	22.15	130.00	1260.0	0.09831	0.10270	0.14790	0.09498	0.1582	0.05395	0.7582	1.0170	5.865	112.40	0.006494	0.018930	0.033910	0.015210	0.01356	0.001997	27.32	30.88	186.80	2398.0	0.15120	0.3150	0.53720	0.23880	0.2768	0.07615	0
568	10.26	12.22	65.75	321.6	0.09996	0.07542	0.01923	0.01968	0.1800	0.06569	0.1911	0.5477	1.348	11.88	0.005682	0.013650	0.008496	0.006929	0.01938	0.002371	11.38	15.65	73.23	394.5	0.13430	0.1650	0.08615	0.06696	0.2937	0.07722	1

569 rows × 31 columns

Metrics accessor: same idea, same display API

The metrics accessor exposes methods such as confusion_matrix(), roc_curve(), precision_recall(), and prediction_error(). Each returns a display (e.g. ConfusionMatrixDisplay) with the same interface: plot(), frame(), set_style(), help().

metrics_display = report.metrics.confusion_matrix()
metrics_display.help()

metrics_display.frame()

	true_label	predicted_label	value	threshold	split	estimator	data_source
356	0	0	53	0.527467	None	LogisticRegression	test
357	0	1	1	0.527467	None	LogisticRegression	test
358	1	0	0	0.527467	None	LogisticRegression	test
359	1	1	89	0.527467	None	LogisticRegression	test

Draw the confusion matrix by calling plot():

metrics_display.plot()

Inspection accessor

The inspection accessor exposes model-specific displays (e.g. coefficients() for linear models, impurity_decrease() for trees). These also return Display objects with the same plot(), frame(), set_style(), and help() methods.

inspection_display = report.inspection.coefficients()
inspection_display.plot(select_k=15, sorting_order="descending")

Same API with CrossValidationReport

The same accessors and display API apply to CrossValidationReport. We use the same dataset and model; only the report type changes.

from skore import CrossValidationReport

cv_report = CrossValidationReport(estimator, X, y, splitter=3)

Again: data, metrics, and inspection return displays with plot(), frame(), and set_style().

cv_report.data.analyze().plot(kind="dist", x="mean radius", y="mean texture")

cv_report.metrics.confusion_matrix().plot()

cv_report.inspection.coefficients().plot(select_k=10, sorting_order="descending")

The same accessors and display API apply to ComparisonReport (metrics and inspection; no data accessor when comparing reports).

Summary

• Reports (Estimator, CrossValidation, Comparison) use the same accessor layout: report.data, report.metrics, report.inspection (where applicable).

• Accessor methods that produce figures or tables return Display objects.

• Displays share a single, predictable API:

  • plot(**kwargs) — render the visualization

  • frame(**kwargs) — return the data as a pandas.DataFrame

  • set_style(policy=..., **kwargs) — customize appearance

  • help() — show available options

This consistency makes it easy to switch between report types and to reuse the same workflow across data, metrics, and inspection.