DESIGNING EVALUATIONS (EVALS)

LEARNING GOALS By the end of this chapter you will be able to explain what an eval is, build a dataset with golden reference answers, distinguish task success metrics from step-level metrics, and design tests that probe specific failure modes on purpose. WHAT IS AN EVAL? In Chapter 31 you learned that an evaluation harness is the machine that runs tests. An eval, short for evaluation, is one of the tests, or a whole collection of them, that the machine runs. More formally, evals are structured tests of an AI system's quality. The word structured is important. Asking a friend to try your agent and tell you what they think is feedback, but it is not an eval. An eval has a defined input, a defined way of judging the result, and a defined pass or fail rule, so that anyone can repeat it and get comparable results. A school exam is a good analogy. An exam has fixed questions, an answer key, and a grading scheme. Two teachers grading the same exam should reach nearly the same score. Good evals aim for that same repeatability.