marin-community/Atlantic-CC-Sample

Atlantic-CC Sample

75,330 openly licensed web documents (~281M tokens) from Atlantic-CC. Every document comes from a page that states a Creative Commons licence (CC BY, CC BY-SA, or CC0) for its own text, and every row records which one.

This is a sample taken from 16,032 Common Crawl WARC files spanning 131 crawls (2013–2026).

Documents	75,330
Tokens (Llama 3)	~281M
Characters	1.19B
Parquet size	483 MB, 2 files
Crawls	131 (2013–2026)
Licences	CC BY-SA, CC BY, CC0

Quick start

from datasets import load_dataset

ds = load_dataset("marin-community/Atlantic-CC-Sample", split="train")
print(ds[0]["url"], ds[0]["license"])

# Only CC BY (attribution, no share-alike)
cc_by = ds.filter(lambda row: row["license"] == "CC-BY")

What's inside

source	docs	share of tokens	licence
Wikimedia projects (Wikipedia, Wikisource, Wikiquote, Wikibooks, Wikivoyage, …)	35,048	56%	CC BY-SA (197 CC BY)
Stack Exchange network (Stack Overflow, Super User, Ask Ubuntu, MathOverflow, …)	34,578	25%	CC BY-SA
Open-access journal articles (MDPI, BioMed Central, Hindawi, …)	5,704	19%	CC BY (5,688), CC BY-SA (14), CC0 (2)

license	docs	share of characters
CC-BY-SA	69,443	80.5%
CC-BY	5,885	19.4%
CC0	2	<0.01%

How documents were selected

Atlantic-CC is a filtered, deduplicated, and decontaminated Common Crawl corpus. This subset keeps only the Atlantic-CC documents whose licence could be confirmed from the original page. Released text no longer has page footers, so each document's original HTML was re-read from Common Crawl and checked against these rules:

1. The HTML must mark a CC BY, CC BY-SA, CC0, or Public Domain Mark licence in a rel="license" link or a licence <meta> tag.
2. Then one of the following must hold:
   • the page is on a host whose text licence is known: Wikimedia projects or the Stack Exchange network;
   • the page has an article-level statement, such as "This is an open access article distributed under the terms of the Creative Commons Attribution License". Site-wide footers alone are not enough.
3. The page is rejected if any of these apply:
   • its licence text carves content out ("unless otherwise noted", "third-party material");
   • it mentions NonCommercial or NoDerivatives terms;
   • it says "all rights reserved";
   • the licence appears only in JSON-LD or inside a form;
   • the page is a library-proxy mirror.

Using the data

Each document carries its own licence, and reusing a document means following it:

• CC BY: credit the source. Use url to identify it.
• CC BY-SA: credit the source, and share adapted material under the same licence.
• CC0: no conditions.

license holds the most restrictive licence the page links. A page linking both CC BY and CC BY-SA is labelled CC-BY-SA. Licence labels come from the pages themselves; this is not legal advice.

Fields

field	type	description
record_id	string	WARC-Record-ID of the source record
url	string	source page URL
crawl	string	crawl ID (e.g. CC-MAIN-2025-08)
warc_path	string	source WARC file, relative to the Common Crawl bucket
text	string	extracted main text
fasttext_score	float	quality score, filter stage 1
pooled_prob	float	quality score, filter stage 2
modernbert_prob	float	quality score, filter stage 3 (null if not reached)
exact_copies	int	exact duplicates in the pre-dedup pool (1 = unique)
fuzzy_cluster_size	int	size of the near-duplicate cluster before dedup; 4557202931 marks the giant chained cluster
decon_matches	list<string>	eval tasks the document overlapped; always empty here, since contaminated documents were excluded
license	string	CC-BY, CC-BY-SA, CC0, or PDM

The first eleven fields are identical to Atlantic-CC's, so rows join back to Atlantic-CC on record_id.

Limitations

• Low recall by design. Only about 0.3% of Atlantic-CC documents in the sampled WARCs passed the rules. The subset over-represents Wikipedia and Stack Exchange compared with Atlantic-CC as a whole.
• Licence versions. license records the licence family, not the version. Stack Exchange content is CC BY-SA 2.5, 3.0, or 4.0 depending on when it was posted.
• Mostly English. A few non-English Wikipedia editions appear, mainly Cebuano and Simple English.
• Page-level evidence. Licences are read from page markup and text. A page that misstates its own licence would be mislabelled here too.