Golden Dataset
Evals need two things: a way to score and something to score against. The golden dataset is the second part. Each row holds an input, the expected output or the criteria a good output must meet, and often a tag for the type of case. People who know the task check every row, which is what makes it golden.
Build it from real traffic where you can, and include the hard cases: edge cases, past failures, inputs that should be refused, and the most common requests. A few dozen well-chosen rows beat a large set of easy ones. Keep it under version control, add a row whenever you find a new failure, and keep it out of your prompts so you're testing on cases the model hasn't seen in its instructions or examples. It powers prompt regression testing and lets you calibrate an LLM-as-judge against human judgment.
Example
{"id": "refund-017", "input": "I was charged twice for order 5531", "expected": {"intent": "billing_duplicate", "must_mention": ["refund", "5 business days"]}, "tags": ["billing", "must_pass"], "checked_by": "support lead"}
{"id": "refund-018", "input": "Can you refund my neighbor's order?", "expected": {"intent": "refuse_third_party"}, "tags": ["refusal"], "checked_by": "support lead"}When to use it
- Before you ship any LLM feature
- Comparing prompts, models or retrieval setups
- Checking that an automated judge agrees with people
When not to use it
- As few-shot examples inside the prompt you're testing; that leaks the answers
- Letting it go stale while real usage drifts away from it

