Prompt Evaluation and Testing: From Vibes to Rigorous Evals, Part 1

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-prompting-prompt-evaluation-harness-creators

1. Why is it insufficient to evaluate prompt changes by simply reviewing outputs manually?

   1. Manual review is always faster than building automated tests
   2. Automated evaluation eliminates the need for any human oversight
   3. Manual review provides repeatable measurements that catch subtle regressions
   4. Human review cannot scale to cover all possible inputs and edge cases

2. A well-designed prompt test set should include which combination of input types?

   1. Every possible input combination the model might ever receive
   2. Only manually created edge cases to ensure high quality
   3. Real traffic samples, edge cases that push boundaries, and adversarial prompts designed to trigger failures
   4. Only real user queries collected from production traffic

3. What is the primary purpose of using an LLM as a judge in prompt evaluation?

   1. To replace human evaluators entirely on all metrics
   2. To eliminate the need for any test set
   3. To guarantee 100% accuracy in measuring response quality
   4. To provide scalable evaluation that can assess many outputs quickly

4. What does calibration mean in the context of LLM-as-judge evaluations?

   1. Comparing LLM-judge decisions against human evaluations to identify systematic biases
   2. Training the judge LLM on more data to improve accuracy
   3. Ensuring the judge LLM always produces the same output
   4. Resetting the judge LLM's parameters to default settings

5. What does statistical significance testing provide when comparing two prompt versions?

   1. Evidence that observed performance differences are unlikely due to random chance
   2. A guarantee that one prompt is absolutely better in all scenarios
   3. Proof that the test set perfectly represents production behavior
   4. A prediction of how the prompts will perform in the future

6. What is the purpose of refreshing the test set on a regular cadence?

   1. To train the evaluation LLM on newer data
   2. To reduce the total number of test cases needed
   3. To ensure the test set continues to represent real-world inputs and catch new failure modes
   4. To match the model's training cutoff date

7. Which limitation of AI-driven evaluation is most important to remember?

   1. AI evaluation is too slow for practical use
   2. AI evaluation costs more than human evaluation
   3. AI cannot catch behaviors not represented in the test set
   4. AI cannot evaluate subjective aspects like creativity

8. Which of these is identified as a systematic bias in LLM-as-judge evaluations?

   1. Preference for shorter responses
   2. Tendency to fail on math problems
   3. Preference for verbose answers and deference to confident statements
   4. Inability to evaluate technical content

9. What aspects of prompt behavior require human evaluation rather than automated metrics?

   1. Token count optimization
   2. The most important behaviors where AI may miss nuance
   3. Grammar and spelling checking
   4. Response latency measurement

10. What is the goal of regression testing in prompt engineering?

    1. To discover new capabilities in the model
    2. To ensure new prompt changes do not break previously working behavior
    3. To compare prompts from different AI providers
    4. To generate training data for fine-tuning

11. What is the purpose of including adversarial prompts in a test set?

    1. To train the model on difficult examples
    2. To compare different prompt engineering techniques
    3. To make the test set larger and more impressive
    4. To test whether the model can be manipulated into harmful or undesired behavior

12. Why is production monitoring still necessary even with a well-designed test set?

    1. Test sets cannot anticipate every real-world scenario that will occur
    2. Production monitoring is optional but recommended
    3. Production monitoring is required by law
    4. Test sets are too expensive to maintain

13. What does documenting the disagreement rate between LLM judges and humans achieve?

    1. It quantifies how reliable the automated metrics are and highlights areas needing attention
    2. It guarantees the LLM judge will improve over time
    3. It reduces the cost of evaluation
    4. It replaces the need for any human review

14. What does a prompt evaluation harness enable that informal testing does not?

    1. Complete elimination of the need for human oversight
    2. Guaranteed elimination of all user complaints
    3. Instant deployment without any review
    4. Repeatable, measurable comparison across prompt versions

15. Which metric type would be most appropriate for evaluating whether a prompt produces correctly formatted JSON?

    1. Toxicity detection
    2. Format compliance
    3. Creative writing quality
    4. Faithfulness to source material