Natural Questions

Evaluates QA systems on a two-stage pipeline: first retrieving relevant Wikipedia passages from 5.9M articles, then extracting answers from those passages. Unlike single-stage QA benchmarks, Natural Questions forces models to solve both information retrieval (finding the right document/passage) and reading comprehension (extracting the answer) in sequence, measuring end-to-end open-domain QA performance with 307,373 real Google Search queries paired with gold Wikipedia articles and human-annotated answers.

Dataset contains 307,373 naturally-occurring questions extracted from anonymized Google Search query logs, preserving the distribution and phrasing of actual user information needs rather than synthetic or crowdsourced questions. Questions span diverse topics, question types (factual, definitional, numerical), and difficulty levels, with natural language variation (typos, fragments, colloquialisms) that synthetic datasets cannot capture. This grounds evaluation in real user behavior and search intent patterns.

Each question is annotated with two complementary answer types: long answers (paragraph-level passages from Wikipedia, marked with start/end character offsets) and short answers (entity-level spans, marked with token indices). Annotators identify both levels from the same Wikipedia article, or mark the question as unanswerable if no answer exists. This dual annotation enables evaluation of both passage-level retrieval quality (can the system find the right paragraph?) and fine-grained answer extraction (can it identify the exact entity or phrase?).

Annotators explicitly label each question as answerable or unanswerable based on whether a valid answer exists in the paired Wikipedia article. Unanswerable questions are not simply omitted — they are included in the benchmark with explicit labels, forcing QA systems to learn to recognize when no answer exists rather than always attempting extraction. This tests a critical capability for production systems: rejecting questions outside the knowledge base rather than hallucinating answers.

Benchmark includes the full 5.9M Wikipedia article corpus (2018 snapshot) as the retrieval target, requiring systems to rank relevant passages above irrelevant ones. Evaluation measures retrieval performance independently of answer extraction — systems are scored on whether they retrieve the correct Wikipedia article and passage before attempting to extract the answer. This decouples retrieval quality from extraction quality, enabling diagnosis of pipeline failures.

Multiple annotators independently annotate each question with long and short answers, enabling measurement of inter-annotator agreement (IAA) and identification of ambiguous or difficult questions. Benchmark includes agreement metrics (e.g., F1 agreement between annotators) for each question, allowing researchers to filter by agreement level or analyze systematic disagreement patterns. This provides insight into question difficulty and annotation quality.

Benchmark enables computation of separate evaluation metrics for retrieval and extraction stages: retrieval metrics (recall@k, MRR) measure whether the correct Wikipedia article is ranked highly, while extraction metrics (F1, exact match) measure whether the answer span is correctly identified. Pipeline metrics (end-to-end F1) measure overall QA performance. This modular evaluation approach allows diagnosis of failures at each stage and comparison of different architectural choices.

Natural Questions spans diverse Wikipedia article categories (science, history, biography, geography, etc.), enabling evaluation of QA system generalization across domains. Questions are paired with articles from different Wikipedia sections, testing whether systems can handle domain-specific terminology, article structures, and information patterns. This provides insight into cross-domain robustness beyond single-domain benchmarks.