AI Engineering · LLM applications that ship
How do you build a golden evaluation set?
Short answer
Start from real user queries, cover common and hard cases, write explicit expected behaviour, and version the set. Review disagreements until the labels are consistent enough to trust.
A golden set is a small, curated collection of inputs with agreed expected outputs or grading criteria. It is not a dump of production logs. The goal is to represent the decisions the feature must get right, including the cases that have failed before.
- Sample real queries across intents, lengths and user types.
- Include edge cases: empty input, ambiguous questions, out-of-scope requests, adversarial prompts.
- Define the expected answer or rubric before looking at model output, to avoid grading toward whatever the model does.
- Version the dataset and record why each case exists.
- Review a sample by hand and measure agreement between reviewers.
Follow-ups you should expect
Interviewers use these to check whether the first answer was understanding or memorisation.
- 1How would you handle a case where two reviewers disagree?
- 2When should you retire an eval case?
- 3How do you stop the set from overfitting to one model?
Say it in the interview
I start with real queries rather than invented examples, then deliberately add the hard cases: ambiguous requests, missing context, out-of-scope questions and past failures. Each case gets an expected behaviour or rubric written before I see model output. I version the set and grow it from production incidents. Fifty sharp cases with clear labels are more useful than a huge unlabelled dump.
Use this as a model, not a script. Saying it in your own words is what makes it survive follow-up pressure.
Related questions
Reading is step one
Now say it out loud.
Understanding an answer and delivering it under pressure are different skills. Rehearse this topic in a live AI Engineering mock interview and get scored feedback on depth, clarity and confidence.