A JAX library for evaluating and improving the local calibration of binary probabilistic classifiers. KiTE helps determine whether a model is locally calibrated, identify the regions where its predictions are over- or under-confident, and learn a recalibration correction when needed.
Test
Run a permutation hypothesis test of whether a model is locally calibrated.
Diagnose
Locate calibration bias across the feature space instead of relying only on a global average.
Recalibrate
Fit a lightweight correction that improves local calibration while preserving the original model's signal.
Install from GitHub
pip install git+https://github.com/ritwikvashistha/kernel_calibration.git