DESIGNING EVALUATIONS (EVALS)

BUILDING DATASETS AND GOLDEN REFERENCE ANSWERS A dataset, in this book, is simply a collection of eval cases kept together. Building a good dataset is one of the most valuable things a harness engineer does, and it is mostly careful, patient human work. What is a golden reference answer? A golden reference answer is a pre-verified, trusted correct answer that you store beside a test input. The word golden signals that this answer is treated as the standard of truth. When the agent answers, its response is compared to the golden answer. Suppose the test question is "What is our refund policy for items over 90 days old?" The golden answer is the exact, carefully verified policy text, written by someone who checked it against the real policy document. If the agent's answer says that such items can be refunded within 120 days, the comparison reveals the mistake. Where do golden answers come from? There are four common sources. First, subject experts write them, such as a senior support worker writing the right resolution for a ticket. Second, they are extracted from authoritative documents, such as a policy manual. Third, they are computed by trusted code, such as the exact total of an invoice. Fourth, they are taken from past cases that humans have already reviewed and approved. A warning is needed here. A golden answer is only as good as the person who wrote it. If the golden answer itself is wrong, the eval will punish a correct agent and reward a wrong one. Therefore golden answers should be reviewed by a second person, and when the policy changes, the golden answers must be updated.