Hugging Face's logo

Datasets:
obadakr
/
commonswire

Modalities:
Text
Formats:
parquet
Languages:
English
Size:
100K - 1M
Tags:
news
journalism
fact-checking
creative-commons
communication-research
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
commonswire / README.md
obadakr's picture
obadakr
CommonsWire 0.2.1 (#1)
f12d53b
preview code
|
Raw History Blame Contribute Delete
12.6 kB
metadata 
pretty_name: CommonsWire
language:
  - en
license: other
license_name: per-row-open-licenses
license_link: https://huggingface.co/datasets/obadakr/commonswire/blob/main/LICENSES.md
tags:
  - news
  - journalism
  - fact-checking
  - creative-commons
  - communication-research
size_categories:
  - 100K<n<1M
configs:
  - config_name: by
    data_files:
      - split: train
        path: data/by/*.parquet
    default: true
  - config_name: by-sa
    data_files:
      - split: train
        path: data/by-sa/*.parquet
  - config_name: pd
    data_files:
      - split: train
        path: data/pd/*.parquet

CommonsWire 0.2.1

CommonsWire is a full-text corpus of news articles whose publishers allow their text to be shared again. Each row is one article: its body text as published, its metadata, its license, and the credit line its publisher asks for.

Version 0.2.1 holds 149,152 articles (55.2 million words) from 7 sources, published from 2004-10-27 to 2026-11-14 and collected up to 2026-11-15.

  • Curated by: Obada Kraishan, College of Media and Communication, Texas Tech University.
  • License: each article keeps its publisher's license (public domain, CC BY or CC BY-SA), on its row; see LICENSES.md.
  • Contact: the issue tracker.

CommonsWire is not affiliated with, endorsed by or sponsored by any of the newsrooms whose articles it holds.

What it is, and what it is not

  • It is each source's own articles, as far as its API or dump gives them, with every article's license on its row. More sources join in later versions.
  • It is not every news outlet. Most newsrooms do not allow their text to be shared again, and they are not here. Results from CommonsWire describe these sources, not "the news".
  • It is small going forward. Over the last 90 days the sources published about 6.2 new articles a day: 360info 0.5, The Public Record 0.6, Balkan Diskurs 0.2, Factly 2.3, Alt News 1.2, Global Voices 1.5. Fact-checks are most of that flow (about 56%).
  • Global Voices is most of it (101,564 of 149,152 rows, 68%): results over the whole corpus mostly describe it.
  • Wikinews is a closed archive. English Wikinews stopped publishing on 2026-05-04.
  • It is English only.
  • It has no coded variables yet (topic, frame, types of sources quoted, tone). They are planned for a later version.

This update

Version 0.2.1 updates 0.2.0 with what the sources published, edited or removed in the 30 days up to 2026-11-15: 5 new rows, 2 changed, 0 removed at their source, and 0 taken down on request.

Changes since 0.1.0

  • A new source: Global Voices (101,564 rows, CC BY 3.0, 2004 to 2026), an international citizen-media newsroom; its writers' names come from its own story listing, which names every writer of a story, and its translators apart.
  • A new field, translators: the translators the source names, apart from the writers in authors (4,963 rows); the credit line names them ("..., by A, translated by B, ...").
  • Rows credited only to an account are held back (264): an account such as "Guest Contributor" names no person.
  • Near-duplicates across sources: 1 pair in this version.

The sources

Source Kind License Rows Published Credit line template
Wikinews (English) news CC BY 2.5 (18,466); CC BY 4.0 (231); Public domain (1,954) 20,651 2004-11-13 to 2026-05-03 {title}, Wikinews ({url}), {license}
360info news CC BY 4.0 2,400 2021-11-22 to 2026-11-12 {title}, written by {authors}. Originally published under Creative Commons by {source} ({url}), {license}
The Public Record news CC BY-SA 4.0 2,751 2013-06-28 to 2026-11-14 {title}, by {authors}, {source} ({url}), {license}
Balkan Diskurs news CC BY 3.0 879 2014-06-04 to 2026-11-13 {title}, by {authors}, originally published by {source} ({url}), {license}
Factly fact-check CC BY 4.0 15,103 2014-12-14 to 2026-11-13 {title}, by {authors}, {source} ({url}), {license}
Alt News fact-check CC BY 3.0 5,804 2016-04-10 to 2026-11-14 {title}, by {authors}, {source} ({url}), {license}
Global Voices news CC BY 3.0 101,564 2004-10-27 to 2026-11-14 This story by {authors} originally appeared on {source}: {title} ({url}), {license}

{authors} is the byline as the page gives it; where an article names no person and the source's own byline is the source itself, the credit line names the source. When the source names translators (Global Voices, for 4,963 rows), {authors} is followed by ", translated by" and their names; the writers stay in authors, and the translators are in translators. 275 more rows are held back because the source asks for the author's name: 264 are credited only to an account such as "Guest Contributor", which names no person; 11 name no author on the article's page or in the source's story listing.

Every source in the registry (included, pending and excluded), with the pages read and the words quoted as evidence, is in sources.csv.

Configurations and licenses

Each row carries its own license, license_url and credit_line. The rows are split into one configuration per license class, so you can load only what your use allows:

Configuration Licenses Rows
by Creative Commons Attribution (CC BY 2.5, 3.0 and 4.0) 144,447
by-sa Creative Commons Attribution-ShareAlike (CC BY-SA 4.0) 2,751
pd public domain 1,954 
from datasets import load_dataset

by = load_dataset("obadakr/commonswire", "by", split="train")  # CC BY
by_sa = load_dataset("obadakr/commonswire", "by-sa", split="train")  # CC BY-SA
pd = load_dataset("obadakr/commonswire", "pd", split="train")  # public domain

What you must do (this is a summary, not legal advice; LICENSES.md has more):

  • CC BY (by): keep each article's credit_line with it. It names the creator and the source, the title, the link and the license. Say if you changed the text.
  • CC BY-SA (by-sa): the same, and share anything you adapt from these texts (a translation, an edited version) under CC BY-SA.
  • Public domain (pd): nothing is required. The credit line is kept as a courtesy. Public-domain status under US law may not hold in every country.
  • In every case, do not suggest that a newsroom endorses your work, and drop articles that a later version marks as removed.

Fields

Field Meaning
id cw_ and 16 hex digits of the SHA-256 of the canonical URL; stable across versions
source_id, kind the source, as in sources.csv, and its kind (news, fact-check)
title the headline as published
authors the bylines as published (empty when the article names no person)
translators the translators the source names, apart from the writers (empty when it names none)
published_at, updated_at UTC, as the source gives them
retrieved_at UTC, when CommonsWire fetched the article
language en
section the source's own section or category, if any
url the article at its source
license, license_url the article's license and a link to its legal text
credit_line the attribution to keep with the text
text the body text, words unchanged; paragraphs separated by a blank line
content_hash SHA-256 of text, to spot changed articles between versions
status live, or changed if the source edited the text after it was first collected
near_duplicates ids of very similar articles in other sources (5-gram Jaccard 0.6 or more)

index.csv lists every row's id, source, configuration, URL, license, content hash and publication time, without text. manifest.json gives the row counts and the size and SHA-256 of every file.

How it was collected

  • Rights first. A source enters only when its own pages give a license or public-domain statement that allows sharing the full text unchanged, do not forbid bulk or automatic reuse, and do not forbid automated collection, and when its own articles can be told apart from material it republishes. The evidence (page, date, exact words) is in sources.csv.
  • Third-party material is left out. Wire copy, partner stories, syndicated pieces and posts under someone else's rights are dropped; 2,476 were dropped in this version.
  • Text unchanged. The body text comes from each source's API (Wikinews: from the Wikimedia dump, with the wiki markup removed by rule). Only whitespace inside paragraphs is normalized. Photos, captions, video, audio and embedded posts are not collected; they are often separate works the source does not license.
  • Bylines. From the source's API; where the API gives none and the source asks for the author's name, from the byline on the article's own page or in the source's own story listing (Global Voices), never from the body text.
  • Duplicates. Exact duplicates (the same URL or the same text) are dropped: 107 in this version. Near-duplicates across sources are flagged in near_duplicates, not dropped: 1 pair in this version.
  • Politely. One request at a time, at least 2 seconds apart, robots.txt and crawl-delays honored, and a user agent that names the project. Blocks are never worked around.

Removals and corrections

  • At the source. Each update asks every source that still publishes for the articles it published or edited in the last 30 days. An article edited at its source, however old, gets its new text and status changed in the next version; an article of the last 30 days that was removed at its source is dropped from the next version. Removals of older articles are not checked yet.
  • On request. Publishers, writers and people named in an article can ask for a removal or a correction with the takedown and correction form. Requests are answered within 14 days. An article taken down goes on a list of ids (data/takedowns.csv in the code repository, with no URL or name) and is left out of every later version.
  • Your copy. If you keep a copy of CommonsWire, please apply removals to it too: index.csv of the latest version lists every article still in it.

Limitations

  • The sources are a small, uneven set: an international citizen-media newsroom, two fact-checking sites in India, a closed citizen-journalism wiki, a local newsroom in Hamilton (Canada), an open-access research newswire, and an outlet in Bosnia and Herzegovina. Their mix of countries and topics is not that of mainstream news.
  • 33% of the Global Voices rows (33,123) have fewer than 200 characters of text: short posts of a sentence or two, most of them with a link, from its first years (99% of them published before 2013). The median Global Voices row has 50 words. Filter by length if you need full stories.
  • Wikinews text is converted from wiki markup by rule; a few articles keep a stray bar, bracket or brace where the wikitext itself has a slip.
  • Fact-check articles and Global Voices stories quote and describe posts from social media and blogs. The embedded posts are removed, but the quoted words stay in the text.
  • authors is filled for 128,367 of 149,152 rows. Wikinews articles carry no bylines, and some articles are credited to the source itself.
  • News articles name people. CommonsWire adds nothing to what was published, infers nothing about anyone (no gender, ethnicity or other attribute), and offers no search by person.

Contact

Questions and errors: the issue tracker. Removals and corrections: the takedown and correction form.

Citation

Please cite the version you used, and credit the newsrooms as each row's credit line asks.

@misc{commonswire_0_2_1,
  author = {Kraishan, Obada},
  title = {CommonsWire: a full-text corpus of openly licensed news},
  version = {0.2.1},
  year = {2026},
  howpublished = {Hugging Face},
  url = {https://huggingface.co/datasets/obadakr/commonswire}
}

The code and the source registry (no article text) are archived on Zenodo: https://doi.org/10.5281/zenodo.24231835 (all versions).

Version 0.2.1, collected up to 2026-11-15. Code: https://github.com/obadaKraishan/commonswire (Apache-2.0).