Building an evaluation set

Without evals, every change is a guess

An evaluation set is a fixed list of inputs with expected outcomes. Twenty carefully chosen cases beat two hundred random ones, provided they include the failures you actually saw in production.

Correctness has to be defined before you measure. Exact match works for extraction, a rubric works for prose, and a model-graded score works for anything in between — as long as the grader is itself checked against human judgement.

Run the whole set on every prompt change. A prompt tweak that fixes one case and breaks three is extremely common and invisible without the set.