← Table of Contents · Page 340 of 868

DESIGNING EVALUATIONS (EVALS)

Part 12 of 12

CHECK YOUR UNDERSTANDING • What is a golden reference answer, and why must it be maintained? ◦ Answer: It is a trusted correct answer stored with a test input, and it must be updated when policies or facts change so it does not punish correct behavior. • Give one thing a step-level metric can reveal that a task success metric cannot. ◦ Answer: It can reveal that a run succeeded through a risky or wasteful path, such as using a forbidden tool. • Why include adversarial cases in a dataset? ◦ Answer: To test robustness against abuse and learn about weaknesses in safe conditions rather than in production. • What is dataset leakage? ◦ Answer: When test answers are accidentally available to the agent, making the agent look better than it really is. • Why run a failure-mode probe more than once? ◦ Answer: Because model behavior can vary between runs, so repeated runs show how often the failure occurs.
Advertisement
Read Aloud:
0:00 / 0:00