What is model calibration? Probabilities should match observed frequencies
Explain probability calibration versus accuracy, reliability groups, held-out data and verbal confidence.
Understand the concept
Calibration compares predicted probabilities with observed frequencies, such as roughly eighty percent outcomes in an eighty-percent group. It differs from accuracy and verbal confidence. It needs meaningful scores and independent labels; a percentage written by an LLM is not automatically calibrated confidence.
Reliable probabilities and correct labels are different dimensions
Guo and scikit-learn distinguish calibration from predictive correctness. A classifier can be accurate yet overconfident. Calibration adjusts score-probability relationships rather than repairing every error. Open language answers first need a defined event; subjective confidence is not a universal probability for every statement.
Check probability groups on held-out data
On labeled public classification data, group similar scores and compare observed positive rates. Separate calibration fitting from final validation and disclose small-group instability. No calibration was fitted or plotted here. Do not invent numeric probabilities from words such as “possibly” and treat them as measured reliability.
Recheck calibration when data changes
New populations can invalidate prior relationships. Version models, calibrators and data and recheck with labels. Include labeling and group inspection, not code-line counts. Choose thresholds around consequences rather than reassuring percentages. Report calibration, accuracy and verbal explanations separately; calibration does not remove review.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.