Metrics Reference
Every metric returns a MetricResult(value, ci, n) — never a bare float.
Calibration
| Name |
Key |
Description |
| ECE |
"ece" |
Expected Calibration Error (equal-width bins) |
| Adaptive ECE |
"adaptive_ece" |
ECE with equal-mass bins |
| smECE |
"smece" |
Smooth ECE via kernel regression (recommended) |
| Debiased ECE |
"debiased_ece" |
Bias-corrected ECE² |
| MCE |
"mce" |
Maximum Calibration Error |
| cwECE |
"cwece" |
Classwise ECE (multiclass) |
Scoring
| Name |
Key |
Description |
| Brier |
"brier" |
Brier score with Murphy decomposition |
| NLL |
"nll" |
Negative log-likelihood |
Discrimination
| Name |
Key |
Description |
| AUROC |
"auroc" |
Area Under ROC curve with DeLong CI |
Usage
# Default metrics (binary)
report = rb.evaluate(y, p)
# Specific metrics
report = rb.evaluate(y, p, metrics=["ece", "auroc", "brier"])
# All metrics
report = rb.evaluate(y, p, metrics="all")