doc_qa
doc_qa dataset — 30 document-grounded QA tasks.
Each task provides 3-6 real-world documentation excerpts and a question. The agent must answer using only the provided documents and cite sources.
Difficulty tiers: - easy (10): answer in a single document, straightforward extraction - medium (10): answer requires synthesizing 2-3 documents - hard (10): answer requires reasoning across documents with distractors