NQ320K Reference Collections

This repository provides reproducible reference constructions of the NQ320K dataset used in generative retrieval research.

The name NQ320K is retained for compatibility with the existing literature. The 320K designation refers to the approximate number of query-document pairs associated with the benchmark and should not be interpreted as the number of unique documents.

Different preprocessing and grouping choices can result in substantially different document collections while all being referred to as NQ320K. This repository makes these choices explicit and provides materialized versions of the resulting datasets.

Quick Start

The recommended configuration is NQ320K-HtmlNorm-PageID and is loaded by default:

from datasets import load_dataset

dataset = load_dataset("artefactory/nq320k")

Each configuration now exposes three splits:

documents = dataset["documents"]
train_queries = dataset["train_queries"]
validation_queries = dataset["validation_queries"]

An exact configuration can be selected explicitly:

dataset = load_dataset(
    "artefactory/nq320k",
    "htmlnorm-pageid",
)

For reproducible experiments, we recommend always reporting the configuration name explicitly.

Naming Convention

Configurations follow:

{text-representation}-{grouping-rule}

The text representations are:

• simplified - simplified Natural Questions document representation.
• simplenorm - normalized simplified document representation.
• ncinorm - title-abstract-content representation following the NCI-style preprocessing used in our reconstruction.
• htmlnorm - HTML-aware normalized document representation.

Available grouping rules include:

• pageid - documents are grouped using their resolved Wikipedia Page ID.
• ncititle - documents are grouped using the reconstructed NCI title definition.
• url - documents are grouped using their Wikipedia URL.
• text4k - documents are grouped using a Text4K equivalence rule corresponding to the DSI-style construction.
• none - documents are not grouped further. Config names use the explicit -none suffix, and files are stored under the none directory.

For example:

htmlnorm-pageid

corresponds to the scientific name:

NQ320K-HtmlNorm-PageID

while:

ncinorm-none

corresponds to:

NQ320K-NciNorm-None

Dataset At A Glance

All configurations contain the same query rows:

Split	Rows	Mean characters	Min	Max
train_queries	307,373	48.71	15	100
validation_queries	7,830	49.02	18	100

Lengths are measured as decoded UTF-8 characters in the text field. They are not tokenizer-specific.

Document collection sizes and document lengths vary by text representation and grouping rule. The no-grouping baseline is listed last within each text representation.

HtmlNorm

Hugging Face config	Reference name	Documents	Mean document characters	Min	Max
htmlnorm-pageid	NQ320K-HtmlNorm-PageID	109,229	17,172.00	0	1,715,099
htmlnorm-ncititle	NQ320K-HtmlNorm-NciTitle	109,739	17,203.60	0	1,715,099
htmlnorm-url	NQ320K-HtmlNorm-URL	231,695	23,453.53	0	1,715,099
htmlnorm-text4k	NQ320K-HtmlNorm-Text4K	184,209	21,182.70	0	1,715,099
htmlnorm-none	NQ320K-HtmlNorm-None	315,203	25,126.20	0	1,715,099

NciNorm

Hugging Face config	Reference name	Documents	Mean document characters	Min	Max
ncinorm-pageid	NQ320K-NciNorm-PageID	109,229	25,664.43	4	518,554
ncinorm-ncititle	NQ320K-NciNorm-NciTitle	109,739	25,703.71	4	518,554
ncinorm-url	NQ320K-NciNorm-URL	231,695	35,139.49	4	518,554
ncinorm-text4k	NQ320K-NciNorm-Text4K	195,815	32,384.37	4	518,554
ncinorm-none	NQ320K-NciNorm-None	315,203	37,525.13	4	518,554

SimpleNorm

Hugging Face config	Reference name	Documents	Mean document characters	Min	Max
simplenorm-pageid	NQ320K-SimpleNorm-PageID	109,229	26,210.02	483	518,805
simplenorm-ncititle	NQ320K-SimpleNorm-NciTitle	109,739	26,249.57	483	518,805
simplenorm-url	NQ320K-SimpleNorm-URL	231,695	35,727.15	483	518,805
simplenorm-text4k	NQ320K-SimpleNorm-Text4K	204,643	33,441.72	483	518,805
simplenorm-none	NQ320K-SimpleNorm-None	315,203	38,105.35	483	518,805

Simplified

Hugging Face config	Reference name	Documents	Mean document characters	Min	Max
simplified-pageid	NQ320K-Simplified-PageID	109,229	33,379.93	1,284	672,378
simplified-ncititle	NQ320K-Simplified-NciTitle	109,739	33,438.82	1,284	672,378
simplified-url	NQ320K-Simplified-URL	231,695	44,576.50	1,284	672,378
simplified-text4k	NQ320K-Simplified-Text4K	202,221	41,657.48	1,284	672,378
simplified-none	NQ320K-Simplified-None	315,203	47,561.24	1,284	672,378

Text Representation Examples

The snippets below use the target document for the first validation_queries row, which is the same across all configurations:

{"docid": "0", "text": "what do the 3 dots mean in math"}

The target document is Therefore sign. Snippets are truncated for readability.

Text representation	Example document text
htmlnorm	Therefore sign In logical argument and mathematical proof, the therefore sign is generally used...
ncinorm	Therefore sign ؋ ​ ₳ ​ ฿ ... Uncommon typography ... In logical argument...
simplenorm	Therefore sign - wikipedia Therefore sign Jump to : navigation , search ∴ Therefore sign...
simplified	Therefore sign - wikipedia <H1> Therefore sign </H1> Jump to : navigation...

Configurations

Validation seen/unseen percentages are computed over unique target documents in validation_queries. A validation target document is considered seen when its docid also appears as a target in train_queries for the same configuration.

The tables are grouped by grouping rule so that coverage changes can be compared within the same document grouping. PageID, NciTitle, and URL use metadata-derived grouping keys, so their document counts and seen/unseen rates are stable across text representations.

PageID Grouping

Hugging Face config	Reference name	Documents	Validation target docs	Seen	Unseen
htmlnorm-pageid	NQ320K-HtmlNorm-PageID	109,229	6,926	76.38%	23.62%
ncinorm-pageid	NQ320K-NciNorm-PageID	109,229	6,926	76.38%	23.62%
simplenorm-pageid	NQ320K-SimpleNorm-PageID	109,229	6,926	76.38%	23.62%
simplified-pageid	NQ320K-Simplified-PageID	109,229	6,926	76.38%	23.62%

NciTitle Grouping

Hugging Face config	Reference name	Documents	Validation target docs	Seen	Unseen
htmlnorm-ncititle	NQ320K-HtmlNorm-NciTitle	109,739	6,930	75.28%	24.72%
ncinorm-ncititle	NQ320K-NciNorm-NciTitle	109,739	6,930	75.28%	24.72%
simplenorm-ncititle	NQ320K-SimpleNorm-NciTitle	109,739	6,930	75.28%	24.72%
simplified-ncititle	NQ320K-Simplified-NciTitle	109,739	6,930	75.28%	24.72%

URL Grouping

Hugging Face config	Reference name	Documents	Validation target docs	Seen	Unseen
htmlnorm-url	NQ320K-HtmlNorm-URL	231,695	7,369	25.16%	74.84%
ncinorm-url	NQ320K-NciNorm-URL	231,695	7,369	25.16%	74.84%
simplenorm-url	NQ320K-SimpleNorm-URL	231,695	7,369	25.16%	74.84%
simplified-url	NQ320K-Simplified-URL	231,695	7,369	25.16%	74.84%

Text4K Grouping

Text4K is text-dependent: the grouping key is derived from the normalized text itself rather than from external metadata. Changing the text representation can therefore change which train and validation targets collapse into the same group, so Text4K document counts and seen/unseen rates vary across text representations.

Hugging Face config	Reference name	Documents	Validation target docs	Seen	Unseen
htmlnorm-text4k	NQ320K-HtmlNorm-Text4K	184,209	7,251	51.66%	48.34%
ncinorm-text4k	NQ320K-NciNorm-Text4K	195,815	7,265	36.23%	63.77%
simplenorm-text4k	NQ320K-SimpleNorm-Text4K	204,643	7,286	43.93%	56.07%
simplified-text4k	NQ320K-Simplified-Text4K	202,221	7,271	44.73%	55.27%

No Grouping

The no-grouping configurations are a baseline where documents are not grouped beyond the package's local row construction. They use the none directory on disk and the explicit -none suffix in the Hugging Face config name.

Hugging Face config	Reference name	Documents	Validation target docs	Seen	Unseen
htmlnorm-none	NQ320K-HtmlNorm-None	315,203	7,830	0.00%	100.00%
ncinorm-none	NQ320K-NciNorm-None	315,203	7,830	0.00%	100.00%
simplenorm-none	NQ320K-SimpleNorm-None	315,203	7,830	0.00%	100.00%
simplified-none	NQ320K-Simplified-None	315,203	7,830	0.00%	100.00%

Not all configurations necessarily correspond to a construction previously used in the literature. Some are provided as controlled reference variants.

File Organization

The repository mirrors the two main construction dimensions:

data/
├── simplified/
├── simplenorm/
├── ncinorm/
└── htmlnorm/

Each complete configuration is stored in its own directory:

data/htmlnorm/pageid/
├── documents.jsonl.gz
├── train_queries.jsonl.gz
├── validation_queries.jsonl.gz
└── metadata.json

Configurations with the none grouping suffix use the none directory:

data/htmlnorm/none/
├── documents.jsonl.gz
├── train_queries.jsonl.gz
├── validation_queries.jsonl.gz
└── metadata.json

Files can be inspected and processed without specialized tooling:

gzip -cd data/htmlnorm/pageid/train_queries.jsonl.gz | head
gzip -cd data/htmlnorm/pageid/documents.jsonl.gz | head

Data Format

All files are gzip-compressed JSON Lines (.jsonl.gz). Each row has the same schema:

Column	Type	Description
docid	string	Document identifier within the selected configuration. For query rows, this is the target document ID.
kind	string	Either document or query.
text	string	Document text or query text.
title	string	Document title when available; empty for query rows.
source_ids	list[string]	Original source identifiers represented by the row. Grouped documents can contain multiple source IDs.

The docid namespace is local to each configuration. Do not compare docid values across configurations unless you explicitly account for the configuration's grouping rule.

License and attribution

This dataset is derived from Natural Questions (Kwiatkowski et al., 2019, Natural Questions: a Benchmark for Question Answering Research, TACL), released by Google under the Creative Commons Attribution-ShareAlike 3.0 license. This dataset is distributed under the same license.

Code: artefactory/redsi.

Citation

Citation information will be added with the first stable release of the dataset.

When reporting experiments, please specify the exact configuration used, for example:

We use NQ320K-HtmlNorm-PageID from the NQ320K Reference Collections.