BEIR Natural Questions (orgrctera/beir_nq)

Overview

This release packages the Natural Questions (NQ) slice of the BEIR (Benchmarking IR) benchmark as a flat, table-oriented dataset for passage retrieval evaluation and tooling. The underlying questions come from Google’s Natural Questions corpus: real user queries paired with evidence in Wikipedia. BEIR reformulates NQ as ad-hoc retrieval: each row is a query (question text), and systems must rank a large corpus of Wikipedia passages so that gold-relevant passages (from BEIR qrels) appear at the top.

Natural Questions is a widely used benchmark for open-domain question answering; in BEIR it is used specifically to evaluate retrieval quality (dense/sparse encoders, re-rankers, and full RAG pipelines) under a fixed corpus + queries + qrels protocol. BEIR reports on the order of 3,452 test queries, a corpus of roughly 2.68M passages, and about 1.2 relevant passages per query on average (benchmark statistics vary slightly by preprocessing version).

This Hub dataset is a tabular export aligned with other CTERA benchmark releases: one row per query with expected_output as a JSON string of relevant corpus document IDs (BEIR qrels-style supervision). Passage text is not inlined; join IDs to the BEIR NQ corpus when building an index.

Task

• Task type: Retrieval for Natural Questions (BEIR formulation): given a natural-language question, retrieve the Wikipedia passages that contain the information needed to answer it.
• Input (input): The question (retrieval query), as in the BEIR NQ query set.
• Reference (expected_output): A JSON string listing relevant corpus document IDs with scores (typically binary 1 for relevant), e.g. [{"id": "doc83108", "score": 1}, ...]. Evaluators retrieve against the full BEIR NQ corpus (distributed with BEIR; not duplicated in every row) and score with standard IR metrics (nDCG@k, MRR, Recall@k, etc.).
• Metadata: Original BEIR query_id and split name are preserved.

Background

Natural Questions (source data)

Natural Questions (Kwiatkowski et al., 2019) is a question answering dataset built from real, anonymized Google Search queries. Annotators were shown a question and a Wikipedia page from the top search results; they marked long answers (usually a paragraph) and short answers (one or more entities) when supported by the page, or indicated that no answer appears. The task reflects real user information needs and Wikipedia-grounded evidence—making it a strong testbed for retrieval and reading comprehension.

• Official site & downloads: Google Research — Natural Questions
• Data repository: google-research-datasets/natural-questions

BEIR reformulation

BEIR (Thakur et al., 2021) standardizes many retrieval datasets into a common corpus + queries + qrels layout so that zero-shot and cross-domain comparisons of lexical, sparse, dense, and re-ranking systems are straightforward. For NQ, documents are Wikipedia passages, queries are Natural Questions test queries, and relevance follows BEIR’s qrels derived from the NQ evaluation setup used in the benchmark.

• Code / resources: UKPLab/beir · beir.ai
• Related Hub layouts: BeIR/nq · BeIR/beir-corpus

Data fields

Column	Type	Description
id	string	Stable UUID for this row in this Hub release.
input	string	Query text (the Natural Questions question).
expected_output	string	JSON array of {"id": "<corpus-doc-id>", "score": 1} objects — relevant passage IDs in the BEIR corpus.
metadata.query_id	string	BEIR NQ query identifier.
metadata.split	string	Split name (this release exposes the BEIR test query set).

Splits

Split	Rows
test	3,452
Total	3,452

Examples

Illustrative rows as stored (IDs and text from the dataset).

Example 1 — single relevant passage (test)

• input: who played mama odie in princess and the frog
• expected_output:

[{"id": "doc83108", "score": 1}]

• metadata.query_id: test2405 · metadata.split: test

Example 2 — multiple relevant passages (test)

• input: when did macbook pro 13 inch come out
• expected_output:

[
  {"id": "doc14183", "score": 1},
  {"id": "doc14196", "score": 1}
]

• metadata.query_id: test377 · metadata.split: test

References

Natural Questions (original corpus)

Abstract (TACL 2019): We present the Natural Questions corpus, a question answering dataset. Questions are naturally occurring queries issued to the Google search engine. Annotators are presented with a Wikipedia page from the top 5 search results and a highlighted long answer or short answer if present on the page, or they indicate no answer from this page. We collected 307,445 training examples with single annotations, 7,830 dev examples with 5-way annotations, and 7,842 test examples with 5-way annotations.

• Natural Questions: A Benchmark for Question Answering Research — ACL Anthology Q19-1026
• Google Research publication page

BEIR (benchmark)

Abstract (arXiv:2104.08663): Existing neural information retrieval (IR) models have often been studied in homogeneous and narrow settings, which has considerably limited insights into their out-of-distribution (OOD) generalization capabilities. To address this, and to facilitate researchers to broadly evaluate the effectiveness of their models, we introduce Benchmarking-IR (BEIR), a robust and heterogeneous evaluation benchmark for information retrieval. We leverage a careful selection of 18 publicly available datasets from diverse text retrieval tasks and domains…

• BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models — arXiv:2104.08663
• NeurIPS 2021 Datasets & Benchmarks (OpenReview)

Citation

If you use the Natural Questions data, cite:

@article{kwiatkowski-etal-2019-natural,
  title = {Natural Questions: A Benchmark for Question Answering Research},
  author = {Kwiatkowski, Tom and Palomaki, Jennimaria and Redfield, Olivia and
            Collins, Michael and Parikh, Ankur and Alberti, Chris and Epstein, Danielle and
            Polosukhin, Illia and Devlin, Jacob and Lee, Kenton and Toutanova, Kristina and
            Jones, Llion and Kelcey, Matthew and Chang, Ming-Wei and Dai, Andrew M. and
            Uszkoreit, Jakob and Le, Quoc and Petrov, Slav},
  journal = {Transactions of the Association for Computational Linguistics},
  volume = {7},
  year = {2019},
  address = {Cambridge, MA},
  publisher = {MIT Press},
  url = {https://aclanthology.org/Q19-1026/},
  doi = {10.1162/tacl_a_00276},
  pages = {452--466}
}

If you use the BEIR benchmark formulation, also cite:

@inproceedings{thakur2021beir,
  title={{BEIR}: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models},
  author={Thakur, Nandan and Reimers, Nils and R{\"u}ckl{\'e}, Andreas and
          Srivastava, Abhishek and Gurevych, Iryna},
  booktitle={Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)},
  year={2021},
  url={https://openreview.net/forum?id=wCu6T5xFjeJ}
}

Provenance

Exported for retrieval evaluation (e.g. Langfuse / internal tooling) with Natural Questions as the BEIR sub-benchmark nq. Corpus passage texts are not inlined in each row; join expected_output document IDs to the BEIR NQ corpus when building an index.