Natural Questions

Overview

Natural Questions (NQ) pairs real questions issued to Google search with passages from Wikipedia. Published by Google (Kwiatkowski et al., TACL 2019), it became one of the standard training and evaluation corpora for open-domain question answering and, through that, for dense retrieval — Dense Passage Retriever is trained on it.

It is one of the 18 zero-shot datasets in BEIR.

As reported in Improving Zero-Shot Ranking with Vespa Hybrid Search:

Natural Questions
Queries4,352
Judgments per query~1.2
RelevanceBinary
Query length9.2 words
Document length76.0 words
Document corpus2.68M

The Domain-Shift Pair

NQ’s main role in this vault is as one half of a comparison. Set against MS MARCO:

MS MARCONatural Questions
Query length5.9 words9.2 words
Document length56.6 words76.0 words
Document corpus8.84M2.68M
SourceWeb search resultsWikipedia passages only

The two are close. Both English, both question-like queries over passage-length documents; the lengths differ by a few words. And that gap is enough: a dense retriever trained on NQ performs strongly in-domain and loses to BM25 on MS MARCO zero-shot, with the BEIR leaderboard showing NQ-trained dense models underperforming BM25 across nearly all its datasets.

This is why the pair is useful pedagogically — it shows that domain shift severe enough to break a single-vector model does not look dramatic from the outside. See Zero-Shot Retrieval.

Shallow Judgments

At ~1.2 judgments per query with binary relevance, NQ sits at the opposite extreme from TREC-COVID’s ~493.5 graded judgments. Many genuinely relevant passages are unlabeled, so recall-oriented conclusions understate true quality — the same limitation as MS MARCO, and a reason to read a BEIR average knowing that its constituent datasets measure with very different precision.

Hybrid ranking gains on NQ are correspondingly small: in Improving Zero-Shot Ranking with Vespa Hybrid Search - part two, BM25 scores 0.327, the distilled ColBERT reranker 0.403, and the hybrid 0.404 — one of the few BEIR datasets where the neural component clearly beats BM25 on its own and fusion adds almost nothing on top.

Beyond Retrieval Benchmarks

NQ also serves as a reference point for what trained query distributions look like. Both Agentic Query Workload and Frontier of Search 2026 make the argument that models trained on short, fluent MS MARCO / Natural Questions queries underperform on the long, operator-laden queries that agents actually emit — NQ’s 9.2-word average being the distribution those models learned.

  • MS MARCO — the domain-shift counterpart
  • BEIR — the suite NQ belongs to
  • TREC-COVID — the deep-judgment contrast

Articles

Source