ccnews-ir-simulated-search-queries

An English news-retrieval dataset with 1.84 million simulated search queries paired with positive CC-News-style document targets.

This dataset contains 1,839,547 English query-document pairs derived from the English subset of multilingual CC-News. Each row is designed so that the associated news document is a positive retrieval target for the generated query. The queries were generated with a Qwen3.5-35B-A3B model quantized to NVFP4 and served through OpenAI-compatible vLLM serve endpoints.

The dataset is intended for information retrieval research and engineering over web and news text. It mixes direct fact-seeking queries with broader discovery-oriented queries, and it mixes natural-language semantic questions with more compressed search-style formulations. The goal is to approximate the range of retrieval queries a practical search system may encounter, rather than producing only polished benchmark-style questions.

The source corpus is intfloat/multilingual_cc_news. For this release, only the English source config en is used, and the public subset name in this dataset is english. The published document field is the source-side article text used during generation, built from the news title followed by a newline and the article body text.

Loading the dataset

from datasets import load_dataset

ds = load_dataset(
    "hotchpotch/ccnews-ir-simulated-search-queries",
    "english",
    split="train",
    streaming=True,
)
row = next(iter(ds))
print(row)

Note that categorical columns are stored as Hugging Face ClassLabel, so some loading modes may return integer IDs instead of string names.

License

This dataset is derived from intfloat/multilingual_cc_news. At the time of this release, that upstream dataset card does not declare an explicit SPDX license field in its README front matter. The upstream card states that it is based on CloverSearch/cc-news-mutlilingual.

For licensing and usage terms of the underlying CC News resources, please review the CC News announcement and GitHub repository:

• Common Crawl mailing list announcement
• commoncrawl/news-crawl

Please ensure that your use of this dataset is compatible with the terms of the upstream CC-News-derived resources you rely on.

Acknowledgements

This dataset is derived from intfloat/multilingual_cc_news, which provides multilingual access to CC-News-derived article text through the Hugging Face datasets interface.

Thank you to the maintainers of intfloat/multilingual_cc_news and the upstream CloverSearch/cc-news-mutlilingual resource. That source dataset made it practical to construct a large English news retrieval dataset with positive targets and varied simulated search behavior.

---

Technical report

Dataset structure

The dataset uses the following top-level columns, in this exact order.

Column	Type	Meaning
query	string	Generated search query intended to resemble realistic user search behavior for finding this article or closely related articles.
document	string	Source-side news text used during generation, represented as title + "\n" + maintext.
url	string	Original article URL inherited from the source dataset.
date_publish	string	Original publication timestamp inherited from the source dataset.
query_form	ClassLabel	Whether the output was generated as a search-style query or a semantic natural-language question.
mode	ClassLabel	Main query style, such as direct, partial, tangential, typo-heavy, or keyword-style generation.
asker_level	ClassLabel	Intended phrasing difficulty or linguistic sophistication level.
answer_type	ClassLabel	How directly the document is expected to answer the query.
clean_or_noisy	ClassLabel	Whether the query was generated as a cleaner polished query or as a noisier search-log-like query.
focus_segment	ClassLabel	Which region of the source document the generator was encouraged to focus on.

query is the simulated user search input and document is the positive retrieval target text. url and date_publish are retained so downstream users can trace rows back to the source article record and perform time-aware analysis if needed.

Example row:

Column	Value
query	Is The Cheviot, The Stag and the Black, Black Oil part of the school curriculum?
document	The Cheviot, The Stag and the Black, Black Oil\nThe day had come. I had waited 43 years to see this play. ... On the way out I asked one of the youngsters if they were there because it is on the school curriculum. She replied, yes. ...
url	http://scotsindependent.scot/?p=1480
date_publish	2016-09-10 00:00:00
query_form	semantic_question
mode	direct
asker_level	middle_school
answer_type	direct
clean_or_noisy	clean
focus_segment	all

Released subsets

The following counts describe the current release.

Subset	Source config	Language	Rows
english	en	English	1,839,547

Control semantics

query_form

Value	Meaning
search_query	A search-engine-style query, often shorter and more compressed.
semantic_question	A natural-language question or question-like search input.

clean_or_noisy

Value	Meaning
clean	A clearer, more polished, more natural query.
noisy	A noisier query that may be rougher, typo-bearing, broader, or more compressed.

mode

Value	Meaning
direct	The query asks for information stated directly in the article.
partial	The query focuses on only part of the article and may be less tightly specified.
tangential	The query stays related to the topic but approaches it from a somewhat different angle.
rough	The query is short, rough, fragmented, or search-log-like.
typo	The query intentionally includes misspellings or typographical noise.
keyword	The query is compressed into a keyword-heavy search style.
vague	The query is broader or more underspecified than a direct lookup.

focus_segment

Value	Meaning
all	The whole article may be used as grounding context.
front	Focus on the earlier part of the article.
middle	Focus on the middle part of the article.
back	Focus on the later part of the article.

asker_level

Value	Meaning
middle_school	Simpler and more accessible phrasing.
high_school	Standard everyday phrasing.
college	More precise or somewhat more specialized phrasing.

answer_type

Value	Meaning
direct	The document is expected to directly answer the query.
indirect_hint	The document is expected to provide only a clue or partial support.
interest	The document is relevant to the topic, but may not directly answer the query.

Empirical distribution

The following counts describe the released dataset after filtering.

query_form

Value	Rows	Share
search_query	917,146	49.857166%
semantic_question	922,401	50.142834%

clean_or_noisy

Value	Rows	Share
clean	1,610,085	87.526168%
noisy	229,462	12.473832%

mode

Value	Rows	Share
direct	1,126,280	61.225943%
partial	403,095	21.912732%
tangential	80,710	4.387493%
rough	45,735	2.486210%
typo	80,340	4.367380%
keyword	68,595	3.728907%
vague	34,792	1.891335%

focus_segment

Value	Rows	Share
all	1,286,935	69.959343%
front	183,942	9.999310%
middle	220,748	12.000128%
back	147,922	8.041219%

asker_level

Value	Rows	Share
middle_school	307,012	16.689544%
high_school	611,660	33.250577%
college	920,875	50.059879%

answer_type

Value	Rows	Share
direct	1,313,425	71.399372%
indirect_hint	263,163	14.305859%
interest	262,959	14.294769%

Construction methodology

This section is written so that a senior data scientist or engineer can reproduce a close variant of the dataset design. The resulting generations will not be bitwise identical without the original prompts, seeds, and serving setup, but the methodological recipe is intended to be clear.

Step 1. Define the source article corpus

Start from intfloat/multilingual_cc_news. For this release, only the English config en is used. Each source row contains at least four relevant fields:

• title
• maintext
• url
• date_publish

The public release uses only the train split from that English config.

Step 2. Filter and normalize source rows

Before query generation, convert each source article into a local filtered document row. The generation source text is built as:

• document = title + "\n" + maintext

Rows are discarded if title is empty, if maintext is empty, or if the concatenated text is shorter than 100 characters after normalization. For the English run used here, this preprocessing retained 1,842,875 source rows.

Each retained row also preserves:

• row_index
• url
• date_publish

The stable generation key is the original row_index from the filtered local dataset.

Step 3. Assign query-generation controls

Each retained row is assigned a deterministic control bundle from its stable row index together with a fixed control seed. The generation controls cover:

• query_form
• clean_or_noisy
• mode
• focus_segment
• asker_level
• answer_type

The release mixes two broad surface forms:

• search_query
• semantic_question

Within those forms, the generator further varies directness, noisiness, target answerability, linguistic sophistication, and which region of the article should be emphasized.

Step 4. Generate search queries with a vLLM-served model

The generation run used a Qwen3.5-35B-A3B model quantized to NVFP4 through OpenAI-compatible vLLM serve endpoints. The prompts asked the model to generate one user-like retrieval query grounded in the article text and constrained by the selected control bundle.

The generation setup used temperature-based sampling rather than greedy decoding. For each successful generation, the raw row JSON stored:

• the generated query
• the source text
• source metadata such as url and date_publish
• control values such as query_form, mode, and answer_type
• generation bookkeeping including ok and error

Step 5. Filter completed generations for public release

The public dataset is not a direct dump of all row JSON files. Instead, only rows satisfying all of the following conditions are exported:

• ok == true
• error == null
• query is non-empty
• text is non-empty
• all categorical control values are valid members of the released schema

For the English run, 1,842,875 generated row JSON files were scanned and 1,839,547 rows were retained for the public dataset. This corresponds to a retention rate of approximately 99.82% relative to the generated row files considered for export.

Step 6. Export the final IR dataset

Export the retained rows as a flat dataset with one query-document pair per row. The public schema keeps the retrieval text fields first, followed by source metadata and generation controls:

• query
• document
• url
• date_publish
• query_form
• mode
• asker_level
• answer_type
• clean_or_noisy
• focus_segment

The stable internal row_index used during generation is intentionally omitted from the public release.

Author

• Yuichi Tateno (@hotchpotch)