scikit-learn adapter

probcal.sklearn (extra: pip install "probcal[sklearn]", scikit-learn ≥ 1.4) provides two estimators. Neither is imported by import probcal — the core stays numpy-only.

Ending a pipeline with a calibrated column

SklearnCalibrator is score-level: its X is the score itself, one column ((n,) or (n, 1); more columns raise). Use it wherever sklearn expects an estimator, or let transform end a Pipeline:

from probcal import BetaCalibrator
from probcal.sklearn import SklearnCalibrator

est = SklearnCalibrator(BetaCalibrator()).fit(scores, y)
est.predict_proba(scores_new)     # (n, 2)
est.transform(scores_new)         # (n, 1) calibrated column

est.calibrator_.interpret()       # the full probcal audit surface
est.calibrator_.interval_inverse(0.0, 0.02, space="logit")
est.calibrator_.to_json("calibrator.json")

Score columns on the margin scale are declared, not guessed: SklearnCalibrator(input="logit") maps them through expit exactly first.

Cross-validated calibration of a classifier

CalibratedClassifier is the drop-in for sklearn.calibration.CalibratedClassifierCV(ensemble=False):

from sklearn.ensemble import HistGradientBoostingClassifier
from probcal.sklearn import CalibratedClassifier

clf = CalibratedClassifier(
    HistGradientBoostingClassifier(), cv=5, random_state=0
).fit(X_train, y_train)
clf.predict_proba(X_new)

# probcal calibrator protocol, delegated — hand clf straight to treecf:
clf.interval_inverse(0.0, 0.02, space="logit")
clf.fingerprint()

cv="prefit" scores the calibration set with the already-fitted estimator; method="decision_function" maps margins through expit before calibration (the calibrator absorbs the monotone distortion this introduces).

Grid search over the calibration map

The probcal calibrator is a nested estimator, so its parameters are grid-searchable:

from sklearn.model_selection import GridSearchCV

gs = GridSearchCV(
    CalibratedClassifier(model, calibrator=BetaCalibrator()),
    {"calibrator__variant": ["a", "ab", "abm"]},
    scoring="neg_log_loss",
    cv=5,
).fit(X, y)

Against CalibratedClassifierCV

	CalibratedClassifierCV(ensemble=False)	probcal.sklearn.CalibratedClassifier
OOF protocol, one map, full refit	yes	yes (verified equivalent)
Calibration methods	sigmoid, isotonic	all 12 probcal calibrators + CalibratorSelector
Parameter interpretation	—	interpret() on the fitted map
Metric CIs	—	probcal.metrics.evaluate bootstrap
Exact inverse maps (policy → raw threshold)	—	interval_inverse / point_inverse, delegated
Serialization	pickle	versioned JSON + fingerprint() (and pickle)
Multiclass	yes	binary only, by design

Estimator-check compliance

Both estimators run sklearn.utils.estimator_checks.parametrize_with_checks in CI on the pinned minimum (1.4) and the latest release. CalibratedClassifier passes the full corpus except the sample-weight ≡ duplication equivalence, which cannot hold through a CV split whose fold assignment depends on n (sklearn's own CV wrappers share this; declared via expected_failed_checks). SklearnCalibrator's one-column contract is inapplicable to the generic multi-feature checks — those are declared expected failures with the reason stated, the same domain-restriction sklearn special-cases its own IsotonicRegression for; the remaining convention checks run live.