Golden datasets and synthetic eval

Curating immutable evaluation baselines and scaling verification via programmatic test generation.

Why it matters

A reliable evaluation suite requires data rigor. We balance manually curated, edge-case heavy 'golden datasets' built by domain experts with high-volume, synthetically generated question-context pairs (e.g., via evolutions) to maintain cheap, exhaustive coverage across vast document corpora.

A golden set is your versioned, reproducible test suite for retrieval+generation. The best items specify more than the expected answer text — they pin the evidence that must be retrieved.

{
  "id": "refund-enterprise-001",
  "question": "Can an enterprise customer get a $12k refund?",
  "must_retrieve": ["refund-policy", "enterprise-exception"],
  "expected_answer_facts": ["30-day window", "over $10k needs legal approval"],
  "forbidden_claims": ["all refunds are automatic"]
}

Why must_retrieve matters. An item that only checks the answer text can pass by luck even when retrieval silently regressed — the model guessed right this time. Pinning the required evidence ids turns a flaky end-to-end check into a precise one that catches retrieval drift before it becomes a wrong answer.

Blend human + synthetic. Hand-curated items (built by domain experts) cover the tricky edge cases that matter; LLM-generated synthetic items give cheap breadth across the corpus. Humans buy precision on the cases you fear; synthesis buys coverage you couldn't afford manually.

Include adversarial 'should refuse' items. A golden set that only contains answerable questions never tests your refusal behaviour. Add questions with insufficient context whose correct answer is 'I don't know' — that's how you keep honesty from regressing.

A golden set is a flight simulator's scenario library. It's not enough to practise smooth landings — you load engine failures and storms (adversarial cases) on purpose, and you check the pilot used the right instruments (must_retrieve), not just that the plane happened to land.

• Answer-only items. Symptom: pass by luck despite retrieval drift. Fix: pin must_retrieve evidence.
• No adversarial/refusal cases. Symptom: honesty regresses unseen. Fix: add 'should refuse' items.
• Unversioned golden set. Symptom: irreproducible runs. Fix: version it like code.
• Synthetic-only. Symptom: misses real edge cases. Fix: blend with human-curated items.