Calibration
How well predicted confidence matches reality: among predictions made at 70% confidence, about 70% should turn out right.
What people get wrong
A model can rank perfectly and be badly calibrated, or be well calibrated and rank poorly. If a downstream decision uses the probability as a number rather than an order, calibration is the property that matters.
A definition is not an understanding. This is taught properly in Measuring sycophancy, refusals and calibration, part of Alignment & RLHF.
Learn this properlyTerms from the same topic