Part II - Evaluation-based system improvement

An LLM response can follow the requested format and still contain incorrect information or fail to answer the question. Evaluating response quality requires representative test questions and clear criteria for judging the answers.

A useful comparison starts with a definition of success and fixed test cases, including important edge cases. The proposed system is compared with a chosen reference system on the complete set and on meaningful groups of questions. When answers have several valid forms, human reviewers or a separately prompted model use written grading criteria and rules for assigning scores. The release decision must also account for uncertainty in these measurements.

Chapter 4 first establishes evaluation criteria and test cases, then explains how metrics, benchmarks, model judges, human ratings, and uncertainty estimates measure answer quality. Chapter 5 uses that evidence to compare changes, diagnose failures, and decide whether a product meets its release requirements.

The first group covers Sections 4.1 to 4.8: criteria, golden sets, classification, output metrics, benchmarks, model judges, human review, and uncertainty. A group-level arrow leads to Sections 5.1 to 5.6: requirements, baseline comparisons, run records, error analysis, release decisions, and the product case. A separate dashed path takes error analysis back to a new case version rather than changing the cases in a completed comparison.
Figure 1: Part II first establishes criteria, cases, and measurement methods, then uses their results to compare system changes and decide on release. Reviewed failures can inform a later version of the case set. Open full-size map (new tab).

Chapters in this part