Cohen's kappa,
Krippendorff's alpha.
When ground truth comes from human raters, measure the raters first. Agreement below the level you expect from the model means the labels, not the model, are the limiting factor.
Pick the metric
that matches
the decision.
Accuracy, ranking quality and calibration answer different questions. Reading a model report means knowing which question each number was built to answer.
START WITH THE DEFINITION ↓ACCURACY · AUC · BRIER · ECE
01 / DEFINITION
The short answer first.What are LLM evaluation metrics?
LLM evaluation metrics are the numerical summaries of a model's behaviour on a test set. Each one supports a different decision: accuracy says whether the model is competent, ranking metrics say whether a threshold can exist, and calibration metrics say whether the model's confidence can be shown to a person.
02 / THE CORE FOUR
What each number is for.| Metric | Range and direction | Decision it supports | MKC2 vs stock base |
|---|---|---|---|
| Accuracy | 0 to 1, higher better | Is the model competent at the task at all? | 93.17% vs 83.09% |
| ROC AUC / ranking quality | 0.5 is chance, higher better | Can any threshold separate the two classes? | 0.9623 vs 0.7281 |
| Brier score | 0 to 1, lower better | How wrong are the probability estimates overall? | 0.0620 vs 0.1274 |
| Expected calibration error | 0 to 1, lower better | Does a stated confidence mean what it says? | 0.0610 vs 0.1363 |
Accuracy is the easiest to read and the easiest to mislead with. On an imbalanced set, a model that always predicts the majority class can score well while being completely uninformative. Always report the class balance next to it.
ROC AUC measures ordering: the probability that a randomly chosen positive scores above a randomly chosen negative. It is threshold-free, which makes it the right metric when the threshold has not been chosen yet, and it is the first thing to check before anyone designs a routing rule. An AUC near 0.5 means no threshold will work, whatever the accuracy says.
Brier score is the mean squared error of probabilistic predictions. It combines being right with being appropriately confident, which makes it a good single summary and a poor diagnostic: a bad Brier score does not tell you which of the two went wrong.
Expected calibration error buckets predictions by stated confidence and measures the gap between claimed and observed accuracy in each bucket. It is the number to quote when a human will see the model's confidence, and it is sensitive to bucket count, so the bucketing scheme should be published alongside it.
03 / BEYOND THE CORE
Metrics for open-ended and retrieval systems.Cohen's kappa,
Recall@k, nDCG,
context precision.
Generation quality is capped by whether the right passage was retrieved at all. Retrieval metrics isolate that ceiling from the model's reasoning.
Support rate and
citation validity.
The share of generated claims traceable to a retrieved source. A direct measure of fabrication risk, and cheaper to compute than a full rubric.
Two older metrics deserve a caution. BLEU and ROUGE measure n-gram overlap with a reference text. They remain useful for tightly constrained tasks such as translation or templated extraction, and they are close to meaningless for open-ended drafting, where a correct answer that shares no vocabulary with the reference scores near zero. Reporting ROUGE on an open-ended task is usually a sign that no one chose a metric deliberately.
Retrieval metrics are treated in full on retrieval evaluation, and groundedness on AI hallucination detection.
04 / HOW METRICS MISLEAD
Three failure patterns to look for.The aggregate hides the slice.
A strong mean across 45 checks can contain one check at chance level. Per-slice reporting is the only way to find it; averaging is how it survives to production.
The metric outlives its threshold.
A threshold chosen on last year's distribution encodes last year's base rate. When the input mix shifts, the same threshold produces a different false-positive rate without any model change.
The metric became the target.
Any metric optimised against for long enough stops measuring what it was chosen for. Rotating a held-out set is the standard defence.
05 / QUESTIONS
Asked when reading or writing a model report.Which single metric should we report?
+
None. The minimum honest report is three: accuracy for competence, a ranking metric for separability and a calibration metric for whether the confidence is usable. Any one of the three can look acceptable while another is failing.
What is a good ROC AUC?
+
It depends entirely on the task and the base rate, so context matters more than a universal threshold. What can be said generally is that values near 0.5 mean the score carries no ordering information, and that the gap between a fine-tuned and a stock model on the same items is more informative than either absolute value.
What is the difference between Brier score and calibration error?
+
Brier score combines discrimination and calibration into one number, so it drops if either gets worse. Expected calibration error isolates the calibration part: it can be near zero for a model that is uninformative but honestly uncertain. Report both, because neither is interpretable alone.
Do these metrics apply to generative text?
+
They apply wherever the system produces a score or a decision, which includes most review, classification and flagging tasks built on generative models. For free-form drafting, the scoring layer has to produce a judgement first, and the metrics then measure that judgement. See LLM-as-a-judge.
How should confidence intervals be reported?
+
Alongside every figure, derived from the item count. Bootstrapping over the test set is adequate and easy to implement. Without intervals, any comparison of two similar models is unreadable, because the reader cannot tell a real difference from sampling noise.
Report the numbers
that carry a decision.
Privilege AI builds evaluation reports that separate competence, separability and calibration instead of collapsing them into one score.