Calibration Set
What is a calibration set?
A calibration set is a set of samples you manually label so you can see if an LLM-as-a-Judge matches your own judgment. You can't just ask a judge to evaluate another LLM's output and trust its scoring—you have to calibrate the judge against your own judgment first. Your calibration set should have a good mix of inputs that include both samples that include the error and samples that do not.
For example, to calibrate a judge that looks for missed sub-groupings in AI-generated opportunity solution trees, you might collect a set of tree nodes—each one a parent with its children—and manually label each node as having a missed grouping or not.
How do you use a calibration set to score a judge?
Run your judge prompt against the calibration set and compare the judge's scoring to your own manually created labels to see how aligned they are. Two measures matter here: recall, the judge's accuracy when there is an error, and specificity, the judge's accuracy when there is no error. If the judge falls short on either measure, iterate on the judge—simplify or extend the instructions, add examples, try a smarter model—and rescore it against your labels until it aligns with your judgment.
Why does a calibration set need to match production data?
A judge is only calibrated for data that looks like its calibration set. A judge that scores well in calibration can still fall apart against real production data—reporting error rates you know are wrong when you look at actual outputs. Calibration isn't static. When that happens, expand your calibration set with samples that better match what you're seeing in production, then refine the judge prompt until it aligns on the new set.
Related terms:
Learn more:
Last Updated: September 16, 2026