AI evaluation

An evaluation measures behavior or performance against stated tasks and criteria. Define the task and scoring rule before comparing systems, since a metric’s usefulness depends on the task, the cost of different mistakes, and the balance of the data. A benchmark combines test data and a procedure for comparison; a baseline gives the result a reference point. For classification, accuracy counts all correct predictions, precision focuses on predicted positives, and recall focuses on actual positives. F1 combines precision and recall using their harmonic mean. Calibration compares predicted probabilities with observed outcomes. Perplexity expresses average next-token prediction uncertainty on text. These measures describe different properties and need not settle the same question.

• Machine Learning Glossary: Metrics

  Defines baseline as a reference model for comparing performance.

• calibration_curve

  Shows how calibration curves compare predicted probabilities with observed positive frequencies.

• Foundations of Natural Language Processing: Lecture 11, Language Models

  Explains perplexity as the exponentiation of average cross-entropy per token.

Evaluation

Measuring behavior or performance against stated tasks and criteria.

Benchmark

A defined collection of tests and a procedure used to compare performance.

Baseline

A reference approach against which another approach is compared.

Accuracy

The proportion of predictions that are correct under a defined scoring rule.

Precision

The share of predicted positive cases that are actually positive.

Recall

The share of actual positive cases successfully identified.

F1 score

The harmonic mean of precision and recall.

Calibration

How well predicted probabilities match observed outcome frequencies.

Perplexity

A measure of how uncertain a model is about observed tokens.

A useful evaluation starts with an explicit requirement and the conditions where it matters. Translate the requirement into observable outcomes, specify which examples count, define how outputs will be scored, and record the tools and procedures used. Keep the test aligned with the intended context: a dataset that does not represent the people or situations of use can produce misleading assessments. Include examples that target known failure modes, and inspect results across relevant groups when appropriate. If you tune a system repeatedly against a test set, information about that set can leak into development and weaken its role as a final check. Separate development choices from final evaluation where possible, and document the limits of the test.

• Cross-validation: evaluating estimator performance

  Explains how repeated tuning against a test set can leak information into model development.

A measured score is an estimate tied to a particular test, procedure, and set of assumptions. Report uncertainty when the evaluation supports it, and be cautious with simple averages or error bars whose assumptions do not fit the data. Inspect examples behind aggregate results to find recurring failure patterns and cases where the scoring rule may miss the intended requirement. Validity asks whether the evidence supports the interpretation being made: a test can be repeatable yet measure the wrong thing, or perform well on its test while generalizing poorly beyond those conditions. Document test details, limitations, and the intended setting so readers can judge what the result does and does not establish.

• AI RMF Core

  Calls for uncertainty measures, comparison, and documented evaluation, with tests relevant to deployment conditions.

• Measure: NIST AI Risk Management Framework Playbook

  Describes construct validity and limits on generalizing beyond validation conditions.