DESIGNING EVALUATIONS (EVALS)

AVOIDING DATASET MISTAKES Several traps are common. One trap is leakage, which means the test answers accidentally appear somewhere the agent can see them, such as in its instructions or memory, so the agent seems smart when it is merely copying. Another trap is duplication, where many cases are nearly the same, which inflates the score without testing anything new. A third trap is a lopsided dataset, where all cases are easy, and so nothing is ever learned from a high score. A fourth trap is stale cases, where the world changed but the cases did not. TASK SUCCESS METRICS VERSUS STEP-LEVEL METRICS A metric is simply a number or yes-or-no result that summarizes how well something is going. There are two broad families of metrics that you must understand.