What is a golden dataset in LLM evaluation?
A golden dataset is a curated set of inputs paired with verified expected answers or grading criteria, reviewed by people. It is the fixed benchmark you re-run after every change to see whether quality improved or regressed on your real use case.
Explain it simply
A teacher keeps a fixed question paper with a verified answer key to compare batches year after year. The golden dataset is your application's answer key.
Build it from real user questions when possible, supplemented with tricky cases. Each entry has an input and either a reference answer, the source passage that should be retrieved, or a checklist of what a good answer must contain. Humans review it so the labels can be trusted.
Keep it stable so scores are comparable between versions, but add new failures you discover in production. Keep a separate untouched set if you also tune prompts on the first one, otherwise you will overfit to your own test. Size matters less than coverage: 100 careful cases beat 1,000 sloppy ones.
Example
{
"input": "What is the late fee after the due date?",
"expected_source": "fee_policy.pdf p.3",
"must_include": ["Rs 200", "after 7 days"],
"must_not_include": ["waived"]
}What to say in the interview
"A golden dataset is a human-verified benchmark of representative inputs and expected outcomes. I use it to compare versions, grow it with real failures, and hold out a clean subset to avoid overfitting my prompts to it."