ruggsea/infini-news-corpus

INFINI-NEWS Corpus

🔎 Search this corpus online: query it with sub-second full-text search and n-gram counts — in the browser or via a public, keyless REST API, no download required — at infini-news.uni-graz.at (API reference).

A multilingual news corpus extracted from Common Crawl CC-News WARC files. One row per article, with body text extracted via trafilatura, WARC provenance, and derived metadata (publish date, language, topic, byte hashes). Covers Aug 2016 – Apr 2026 (117 crawl months).

This is release v2. It replaces the earlier extraction (v1). What changed:

• 1,372,589,195 articles in 49,748 parquet files (v1: 1,357,027,742 in 48,978 files).
• Language (CommonLingua) and topic (IPTC) labels moved out of the article files into two annotation sets, one file per crawl month, joined on warc_record_id (see Schema). The GlotLID language columns stay inline.
• Row offsets belong to one release: v2 rows go with the v2 FM-index, its metadata store and release/shard_manifest.parquet. Do not mix v1 offsets with v2 files.
• Release metadata (acceptance checks, model revisions, build recipes, row checks) lives in release/.

v1 remains in this repository's history.

At a glance

Articles	1 372 589 195
Distinct hostnames	135 144
Text size	~3.60 trillion characters of extracted body text
Languages — GlotLID (ISO 639-3)	1 214
Languages — CommonLingua (334-class)	331
Files	49 748 article parquet shards, partitioned year=YYYY/month=MM, plus 2 × 117 monthly annotation files
On-disk size	~1.67 TB of article files (zstd parquet)
License	Derived columns CC-BY-4.0; article text subject to publishers' copyright

Articles per year

year	articles	months
2016	7.4 M	Aug–Dec
2017	71.2 M	full
2018	91.5 M	full
2019	129.3 M	full
2020	185.2 M	full
2021	196.9 M	full
2022	208.6 M (peak)	full
2023	193.4 M	full
2024	142.3 M	full
2025	120.1 M	full
2026	26.7 M	Jan–Apr

Top languages

Detected with two independent identifiers — GlotLID (Kargaran et al. 2023), stored inline in the article files, and CommonLingua (Pleias / GSMA 2026), stored in the commonlingua annotation set; counts below are from CommonLingua (334-class, byte-level).

ISO 639-3	articles	share
eng	531.0 M	38.7 %
spa	136.8 M	9.97 %
rus	89.5 M	6.52 %
deu	87.3 M	6.36 %
ita	71.5 M	5.21 %
fra	55.5 M	4.04 %
tur	44.5 M	3.24 %
arb	28.1 M	2.05 %
por	27.1 M	1.97 %
hin	22.4 M	1.63 %
other 321 langs (CommonLingua)	279.0 M	20.3 %

Schema

Article files (data/year=YYYY/month=MM/part-*.parquet):

column	type	description
url	string	full article URL as crawled
url_hostname	string	parsed hostname (e.g. bbc.co.uk)
url_tld	string	parsed top-level domain (e.g. co.uk)
warc_filename	string	source WARC archive file in the CC-News bucket
warc_record_id	string	WARC URN; primary key, unique per article, and the join key for the annotation files
warc_target_uri	string	original WARC-Target-URI header
warc_date	timestamp (UTC)	WARC capture timestamp
warc_payload_digest	string	original payload SHA-1 from the WARC header
http_status	int	HTTP response code at capture time
publish_date	string	publish date parsed from page metadata, when extractable; can be missing or inaccurate
title	string	article title from page metadata
author	string	author byline from page metadata, when present
sitename	string	publisher / site name from page metadata
categories	list<string>	publisher-provided categories
tags	list<string>	publisher-provided tags
description	string	OpenGraph / meta-description string, when present
text	string	extracted article body (trafilatura)
text_xxhash64	uint64	xxHash64 of text (handy as a dedup / equality key)
raw_html_length	int	length of the original HTML payload, bytes
text_length	int	length of the extracted body text, characters
language	string	GlotLID full label (e.g. eng_Latn)
language_iso639_3	string	GlotLID ISO 639-3 code (e.g. eng)
language_script	string	GlotLID ISO 15924 script code (e.g. Latn)
language_score	float	GlotLID confidence ∈ [0, 1]
language_short	string	GlotLID short tag (ISO 639-1 / -3 fallback)
language_short_score	float	GlotLID short-tag confidence
trafilatura_version	string	trafilatura release used for extraction
extraction_version	string	pipeline version tag
extraction_timestamp	timestamp (UTC)	time of body extraction

The crawl year and month are the partition keys in the file path (year=YYYY/month=MM).

Annotation files, one per crawl month, each covering every article of that month (join on warc_record_id):

configuration	path	columns	model revision
commonlingua	annotations/language/year=YYYY/month=MM/part-00000.parquet	warc_record_id, lang (CommonLingua label, 334-class, byte-level), lang_score (confidence ∈ [0, 1]), lang_method (decoder variant)	PleIAs/CommonLingua @ 43fe88d75e94b11283b66daccbbe4a73e7bc1361
iptc	annotations/topic/year=YYYY/month=MM/part-00000.parquet	warc_record_id, iptc_topic (IPTC top-17 Media Topic), iptc_score (softmax probability), iptc_method (classifier version)	classla/multilingual-IPTC-news-topic-classifier @ 4d69cc69bbad47e815337ff70889c1d246b43cff

iptc_topic is one of politics, sport, economy_business_finance, arts_culture, crime_law, human_interest, disaster_accident, health, conflict_war, lifestyle, education, science_tech, society, environment, labour, weather, religion (Kuzman & Ljubešić 2025). Each annotation set has 117 files and 1,372,589,195 rows. CommonLingua labels every article but one, whose formatted classifier input is empty; it keeps an empty label with score zero. CommonLingua reads the title, description and first 1,024 body characters, truncated to 512 UTF-8 bytes.

Quick start

from datasets import load_dataset
ds = load_dataset("ruggsea/infini-news-corpus", "year_2022",
                  split="train", streaming=True)
# Pin a commit for reproducible downloads: add revision="<commit sha>".
# text_xxhash64 is uint64; HF datasets can't auto-cast it to int64.
ds = ds.remove_columns(["text_xxhash64"])
for row in ds.take(3):
    print(row["title"], "|", row["language_iso639_3"], "|", row["url_hostname"])

INSTALL httpfs; LOAD httpfs;
-- Use the hf:// scheme so the part-*.parquet glob resolves (the
-- generic https:// URL forbids globs unless you SET
-- allow_asterisks_in_http_paths = true).
SELECT t.iptc_topic, COUNT(*) AS n
FROM read_parquet('hf://datasets/ruggsea/infini-news-corpus/data/year=2022/month=06/part-*.parquet') a
JOIN read_parquet('hf://datasets/ruggsea/infini-news-corpus/annotations/topic/year=2022/month=06/part-00000.parquet') t USING (warc_record_id)
JOIN read_parquet('hf://datasets/ruggsea/infini-news-corpus/annotations/language/year=2022/month=06/part-00000.parquet') l USING (warc_record_id)
WHERE l.lang = 'eng' AND a.text_length > 500
GROUP BY t.iptc_topic ORDER BY n DESC LIMIT 5;

Companion FM-index

Pre-built FM-index over the same corpus, with sub-10 ms exact substring count over all 1.37 B articles: ruggsea/infini-news-index. The query API serves it; check /api/v2/info for the release it has loaded before using v2 row offsets, and for every retrieved article compare warc_record_id and the string form of text_xxhash64 with the API metadata.

Dataset creation

Curation rationale

A broad, multilingual news corpus that is queryable in milliseconds via the companion FM-index (ruggsea/infini-news-index), with stable provenance back to the original WARCs and lightweight derived metadata (language, topic) shipped alongside, so downstream filtering doesn't require a secondary classifier pass. Together the corpus + index let you do exact substring search and count over 1.37 B articles without scanning the parquet.

Source data

The corpus is built from Common Crawl News (CC-News) WARC files — the weekly archival snapshots that Common Crawl publishes for the news subset of its crawls. WARC (Web ARChive, ISO 28500) is the standard binary container format used by web archives; each WARC file holds a stream of HTTP responses with their full headers and payloads, exactly as the crawler observed them. CC-News covers Aug 2016 — Apr 2026 in this release. Common Crawl re-crawls a curated seed list of news domains plus URLs discovered in its main crawls. The raw WARC files live at commoncrawl/CC-NEWS/; each row in this dataset preserves the exact warc_filename (file the article was read from), warc_target_uri (the URL crawled), warc_date (capture timestamp), and warc_payload_digest (SHA-1 of the original response body). Repeated crawls and syndicated articles remain separate rows, so occurrence frequencies are preserved.

Processing pipeline

1. Stream WARC payloads from the CC-News S3 bucket.
2. Extract body text with trafilatura.
3. Tag languages with GlotLID (Kargaran et al. 2023) during extraction (inline columns), and with CommonLingua (Pleias / GSMA 2026) as a separate annotation set — both labels stored independently, no winner picked.
4. Classify each article into the IPTC top-17 Media Topic taxonomy with the XLM-RoBERTa-large model from Kuzman & Ljubešić 2025 (classla/multilingual-IPTC-news-topic-classifier).
5. Write monthly parquet shards partitioned year=YYYY/month=MM, and pack each annotation set into one file per crawl month.
6. Accept the release with automated checks: extraction provenance verified, extraction failures reconciled (0 remaining), WARC integrity checked. 23 WARCs are truncated at the Common Crawl source; their complete records were kept. The checks, model revisions and build recipes are in release/.

Full code: git.ksolovev.com/infini-news.

Known limitations

News coverage reflects Common Crawl acquisition and publisher availability. Extraction, language prediction and topic prediction can contain errors; the predictions have automated validation, and human quality evaluation is still pending.

Terms of use

The underlying articles were extracted from CommonCrawl CC-News and remain subject to the original publishers' copyright; we redistribute them on the legal basis of research exceptions recognised in most jurisdictions, mirroring the access model used by comparable web-derived corpora. Derived metadata columns (language, topic, extraction provenance) are released under CC-BY-4.0 (see the data license); downstream users are responsible for compliance with copyright law in their jurisdiction.

The companion FM-index at ruggsea/infini-news-index is the open-access query interface.

Operators of news outlets who want their content excluded from future releases can open a discussion on the Hugging Face dataset page citing the affected url_hostname.

FAIR compliance

This release follows the FAIR data principles (Wilkinson et al. 2016):

• Findable. Indexed on the Hugging Face Hub, assigned a persistent DOI 10.57967/hf/8606, and accompanied by this dataset card describing schema, coverage and processing.
• Accessible. Articles are retrievable over HTTPS through the huggingface_hub client, and the companion FM-index is open for queries without downloading the corpus.
• Interoperable. Data are in Apache Parquet with the explicit per-column schema above. Identifiers use standard vocabularies: ISO 639-3 language codes, ISO 15924 script codes, the IPTC top-17 Media Topic taxonomy, and the original WARC URNs.
• Reusable. Provenance is preserved row-by-row (warc_filename, warc_target_uri, warc_payload_digest, warc_date); pipeline versions are captured in trafilatura_version, extraction_version, extraction_timestamp, iptc_method and lang_method, and the release checks in release/. The derived-metadata columns carry CC-BY-4.0; reuse of the article bodies is governed by the Terms of use above.

Citation

@misc{lazzaroni2026infininews,
  author        = {Lazzaroni, Ruggero Marino and Lasser, Jana and Solovev, Kirill},
  title         = {{Infini-News: Efficiently Queryable Access to 1.3 Billion Processed Common Crawl News Articles}},
  year          = {2026},
  eprint        = {2605.18337},
  archivePrefix = {arXiv},
  url           = {https://huggingface.co/datasets/ruggsea/infini-news-corpus}
}