Lesson 3 of 5, 15 min with the exercise
Build a test set
Be able to tell whether a change made things better.
A test set is a list of real inputs with the right answers written down. 30 to 50 cases is plenty to start, as long as they're representative.
- Include the common cases in roughly their real proportions.
- Add the difficult ones you know about.
- Have a person who knows the work agree the right answers.
- Keep it separate from the examples you put in your prompt.
When a mistake turns up in live use, add it to the set. Over time, it becomes a record of everything that's gone wrong.
See it in practice
Spreadsheet format: Case ID, Input, Correct answer, Output, Pass or fail, Error type. For open-ended outputs such as summaries, write what a good answer must include and must not say, then have a person or a separate check score against that list.
Common mistake: Choosing only the cases where you know it works. A test set that always passes tells you nothing.
Try it
Pull 30 past cases into a spreadsheet with an input column and a correct answer column. Use the evaluation scorecard template to score them.
You're done when
- 30 or more real cases
- Someone who knows the work agreed the answers
- The set is separate from your prompt examples