README.md

metadata

license: odc-by
language:
  - en
task_categories:
  - text-classification
tags:
  - common-crawl
  - web
  - human-written
  - pre-chatgpt
  - llm-detection
  - ai-text-detection
  - machine-generated-text-detection
pretty_name: cc-re-2021
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

cc-re-2021: pre-ChatGPT human-written English web text

499,752 documents of clean English, overwhelmingly human-written web text, extracted from the 9 CommonCrawl crawls of 2021, globally deduplicated within the year (and sharing no identical document with the sister year), annotated with WebOrganizer topic × format labels, and selected so the per-format and per-topic row totals are each as uniform as the eligible pool allows (water-filled marginal targets + exact LP; scarce labels contribute their entire pool — see reports/selection_report.json achieved_marginals for the realized totals).

This corpus is the "human" class for training and evaluating LLM-text detectors. Two properties make it suitable:

1. Temporal purity. Every page was crawled in 2021 — before the release of ChatGPT (Nov 2022) and before large-scale LLM text existed on the public web — which rules out chatbot-era AI contamination by crawl date. Pre-LLM machine text (synonym-spinner spam, raw machine translation, template/doorway pages) does exist on the 2021 web; after remediation a manual audit measures its residual at ~5% of rows (95% CI 0.7–9.3; see Known limitations).
2. No LLM in the loop. No LLM ever generated, rewrote, paraphrased, or summarized any text in this dataset. Models are used read-only: language ID, fastText quality scoring, and topic/format classification. The text bytes are what trafilatura extracted from the original 2021 HTML, post-processed only by (a) FineWeb's standard heuristic filters (the C4 rules also drop boilerplate lines — cookie/policy notices, javascript mentions, sub-3-word lines — and strip bracketed citation markers, exactly as in FineWeb), and (b) a deterministic, removal-only line scrubber for page furniture (cookie banners, share widgets, bylines/postmeta, breadcrumbs, pagination, footer notices — 18 documented rules, no model involved; boiler_lines_removed records the per-row count, ~5.6% of rows touched, ~0.06% of characters). No masking or any other rewriting is applied. Some navigation/UI residue still survives in a minority of rows — deterministic scrubbing is precision-first, not exhaustive.

Sister dataset: the corresponding corpus for the other pilot year at G-reen/cc-re-2020 / G-reen/cc-re-2021.

Pipeline

The pipeline composes the published recipes of FineWeb, GneissWeb, Ultra-FineWeb, and WebOrganizer (extraction → filtering → dedup → annotation → balanced selection). Full source code: the ccre pipeline repo (see reports/pipeline_config.yaml in this repo for the exact configuration used).

1. Manifest — enumerate the year's CommonCrawl dumps; sample WARC files uniformly across crawls and segments (seeded, reproducible).
2. Extraction + heuristic filtering (exact FineWeb stage-A recipe, via datatrove): WARC records (not WET) → UT1 adult/URL blocklist → trafilatura (favour_precision=True, no comments) → fastText lid.176 English ≥ 0.65 → Gopher repetition filter → Gopher quality filter → C4 filters (without the terminal-punctuation line rule) → FineWeb quality filters.
3. Cheap junk fences (from GneissWeb, applied as lenient floors): extreme-tokenized-document filter 0.1 < tokens/char < 0.5 (catches mixed-script leakage, encoded junk, and permutation spam that language ID misses); McAlpine-EFLAW readability is stored as metadata only — never used as a filter, to preserve register diversity.
4. Quality floors, not ceilings (Ultra-FineWeb + DCLM): every document is scored by the Ultra-FineWeb-en fastText classifier (with its exact DeepSeek-V2-token preprocessing) and the DCLM fastText. Keep rule: dclm_score ≥ 0.002 (GneissWeb's permissive floor); the UFW score is recorded as metadata only (no UFW floor is applied — a UFW gate would skew the corpus toward educational text). These thresholds are deliberately permissive floors: the goal is diverse, representative human writing, not FineWeb-edu-style "educational" ceilings that would collapse the register distribution a detector must learn.
5. Global per-year MinHash deduplication (datatrove 4-stage MinHash: signatures → buckets → clusters → filter), sha1 64-bit, 14 buckets × 8 hashes over word 5-grams (≈ 0.72 Jaccard threshold), across all of the year's crawls (unlike FineWeb's per-dump dedup — uniqueness matters more than benchmark-optimal mixing for detector data). Duplicate-cluster sizes are retained as metadata.
6. Topic × format annotation (WebOrganizer): the URL-aware TopicClassifier and FormatClassifier (24 labels each, gte-base-en-v1.5 backbone, max_length 512) label every document; softmax confidences are stored.
7. Marginal-uniform selection — the per-FORMAT and per-TOPIC row totals are equalized as far as pool supply allows: water-filled marginal targets per axis (labels whose entire eligible pool is below the common level contribute everything they have; all other labels get an equal share), reconciled at the topic × format cell level by an exact LP (HiGHS) minimizing total L1 deviation from both axes' targets subject to per-cell pool capacities, down to 500,000 rows; within-cell choice is random (seeded). Achieved marginals ship in reports/selection_report.json (achieved_marginals). Note the marginals are as-uniform-as-feasible, not exactly uniform: scarce labels are supply-limited, and residual joint infeasibility (a label's supply concentrated in saturated labels of the other axis) is split across axes by the LP. Before allocation, eligibility filters exclude non-prose and corrupted rows: dropped topics ['Adult'], dropped formats ['Truncated', 'Content Listing', 'Structured Data'], language_score >= 0.925, and tokens_per_char in [0.13, 0.3]. Formats in {'Spam / Ads': 0.01} are deterministically downsampled to the given keep-fraction (the Spam / Ads label was measured ~96% machine-generated junk in a manual audit; a 1% remnant is retained deliberately). Cross-year hygiene: rows whose exact text appears in the other published cc-re year are excluded per cross_year_exclude ({2021: 2020}), so the two years share no identical documents and can serve as leakage-free train/val splits. Seeded and reproducible; see reports/selection_report.json for the full funnel, filter exclusion counts, and per-cell counts.

Test-set decontamination

On 2026-09-19, 248 of the 500,000 selected rows were removed because they overlap the held-out human benchmark rounaksaha12/ai-det-test-human-refined (1,689 FineWeb documents, 2013–2020 crawls), leaving 499,752 rows. Every row was compared with both the benchmark's text_refined and text_original on distinct word 8-grams of the lowercased, NFKC-normalised \w+ token stream, and removed if any of:

• its URL equals a benchmark URL (scheme, www., fragment and trailing slash ignored);
• it shares ≥ 25 distinct 8-grams with a single benchmark document (a verbatim passage of ~32+ words);
• ≥ 5% of its 8-grams occur in a single benchmark document, or ≥ 5% of a benchmark document's 8-grams occur in it.

This is deliberately much stricter than near-duplicate detection: besides re-crawled copies of benchmark pages it removes pages that only share a template, a long quoted passage (scripture, statutes, licences) or a site-wide footer with a benchmark document. After removal, no remaining row shares more than 24 8-grams or 5% containment with any benchmark document. Removed rows, the rule that fired and the matched benchmark doc_id are listed in reports/test_decontamination_removed.csv. The Pipeline description, the statistics below and the other files under reports/ describe the corpus as selected, before this removal (a 0.050% change).

Schema

column	type	description
boiler_lines_removed	int32	Lines removed by the deterministic page-furniture scrubber (0 = untouched)
crawl	string	CommonCrawl dump id, e.g. CC-MAIN-2020-24
date	string	Page fetch time (WARC-Date, ISO 8601)
dclm_score	float32	DCLM fastText (fasttext-oh-eli5) __label__hq probability
eflaw_readability	—
format	string	WebOrganizer format label (24-way)
format_prob	float32	Format classifier softmax confidence
id	string	Stable unique id: <crawl>/<warc-file#record>
language_score	float32	fastText lid.176 P(en); kept only if ≥ 0.65
n_emails	int32	Count of email addresses in text (FineWeb PII regex); text is NOT masked
n_ips	int32	Count of public IPv4 addresses in text (FineWeb PII regex); text is NOT masked
text	string	Extracted clean text (trafilatura; never LLM-generated or rewritten)
tokens_per_char	float32	GneissWeb extreme-token stat: tokens / chars (GPT-2)
topic	string	WebOrganizer topic label (24-way)
topic_prob	float32	Topic classifier softmax confidence
ufw_score	float32	Ultra-FineWeb-en fastText quality score (P(pos))
url	string	Source page URL (WARC-Target-URI)
word_count	int32	Whitespace word count

Statistics

Overview

Statistic	Value
Rows	499,752
Parquet shards	1
Total words	389,643,384
Total text characters	2,278,657,811
Crawls	9
Fetch-date range	2021-01-15T13:41:43Z → 2021-12-09T09:04:59Z

Distributions

Metric	Mean	Median	p5	p95	Min	Max
Words per document	779.3	444	97	2,350	50	97,179
Characters per document	4,557	2,605	562	1.371e+04	224	543,799
Ultra-FineWeb quality score	0.151	0.02891	0.0003666	0.8197	0	1

Rows per crawl

Crawl	Rows	Share
CC-MAIN-2021-04	53,799	10.76%
CC-MAIN-2021-10	54,494	10.90%
CC-MAIN-2021-17	64,380	12.88%
CC-MAIN-2021-21	52,561	10.51%
CC-MAIN-2021-25	48,741	9.75%
CC-MAIN-2021-31	61,034	12.21%
CC-MAIN-2021-39	53,974	10.79%
CC-MAIN-2021-43	62,652	12.53%
CC-MAIN-2021-49	48,365	9.67%

Topics

Topic	Rows	Share
Sports & Fitness	21,788	4.36%
Travel	21,777	4.36%
Health	21,773	4.35%
Entertainment	21,771	4.35%
Crime & Law	21,761	4.35%
Food & Dining	21,746	4.35%
Industrial	21,745	4.35%
Social Life	21,744	4.35%
Games	21,744	4.35%
Religion	21,740	4.35%
Fashion & Beauty	21,739	4.35%
History	21,737	4.35%
Home & Hobbies	21,736	4.35%
Hardware	21,736	4.35%
Science & Tech.	21,733	4.35%
Education & Jobs	21,732	4.35%
Literature	21,726	4.35%
Finance & Business	21,725	4.34%
Art & Design	21,724	4.34%
Transportation	21,720	4.34%
Politics	21,719	4.34%
Software Dev.	21,698	4.34%
Software	21,686	4.34%

Formats

Format	Rows	Share
Listicle	28,832	5.77%
Personal Blog	28,814	5.76%
Audio Transcript	28,811	5.76%
Tutorial	28,801	5.76%
News Article	28,795	5.76%
Nonfiction Writing	28,794	5.76%
About (Org.)	28,789	5.76%
News (Org.)	28,777	5.76%
Academic Writing	28,769	5.75%
Product Page	28,768	5.75%
Creative Writing	28,767	5.75%
Knowledge Article	28,756	5.75%
User Review	28,738	5.75%
Comment Section	28,689	5.74%
About (Pers.)	26,970	5.39%
Customer Support	21,590	4.32%
Q&A Forum	20,971	4.19%
FAQ	15,652	3.13%
Legal Notices	8,876	1.78%
Documentation	2,537	0.51%
Spam / Ads	504	0.10%

Topic × format cells

Non-empty cells: 349 (of 24 × 24 = 576 possible). Largest cell: Literature × Creative Writing (15,772 rows); smallest non-empty cell: Software Dev. × Spam / Ads (1 rows).

Top 30 cells:

Topic	Format	Rows
Literature	Creative Writing	15,772
Food & Dining	Tutorial	9,815
Industrial	Product Page	9,683
Entertainment	User Review	9,258
Art & Design	Product Page	8,943
Travel	Personal Blog	8,889
Politics	News (Org.)	8,465
Science & Tech.	Nonfiction Writing	8,221
Home & Hobbies	Listicle	8,141
Crime & Law	Knowledge Article	7,743
Science & Tech.	Academic Writing	6,977
Education & Jobs	About (Org.)	6,641
Games	Comment Section	6,371
Fashion & Beauty	Personal Blog	6,162
Sports & Fitness	News Article	5,850
History	Knowledge Article	5,749
Politics	Audio Transcript	5,624
Software Dev.	Personal Blog	5,543
Hardware	Product Page	5,497
Travel	Listicle	5,426
Social Life	Nonfiction Writing	5,415
History	Nonfiction Writing	5,365
Hardware	News Article	5,235
Transportation	Comment Section	5,074
Software	Tutorial	5,045
Religion	Knowledge Article	4,861
Games	Personal Blog	4,827
Hardware	Comment Section	4,506
Education & Jobs	About (Pers.)	4,337
Finance & Business	Customer Support	4,215

Known limitations (measured, from manual audits)

Rates from hand-read random samples (n=100–120) and full-corpus scans; details in reports/audit_post_remediation.md:

Limitation	Measured rate	Note
Pre-LLM machine text (spinners, raw MT, doorway/template pages)	~5% of rows (CI 0.7–9.3)	not concentrated in any filterable label
Content-farm / SEO filler (human-typed, formulaic)	~11%	possible home of undetectable pre-2022 neural text
Page-furniture residue (bylines, widgets, notices)	~1 in 4 rows has some	scrubber is precision-first, not exhaustive
Extraction damage (pipe residue, dropped lists, merged words)	2.8% / ~9% / ~10%	inherent to HTML→text; not remediable without re-extraction
Scraped/republished third-party content	~12%	url/date are not reliable provenance for these
Scrubber over-removal (real lines lost)	~0.1% of rows	rows with boiler_lines_removed >= 2 are the risk set
Low-confidence format labels	format_prob < 0.5 on ~10%	filter on format_prob before conditioning on format

Also note: word_count and all scores (language_score, ufw_score, dclm_score, tokens_per_char, eflaw_readability) were computed on the pre-scrub text; rows with boiler_lines_removed > 0 (~5.5%) are slightly shorter than word_count claims.

Provenance, limitations, and bias

• Source: CommonCrawl 2021 crawls, fetched from https://data.commoncrawl.org. Use of the underlying data is subject to the CommonCrawl Terms of Use.
• Extraction bias: text is trafilatura-extracted with favour_precision=True; most boilerplate, navigation, and comments are removed (extraction + C4 line rules + the deterministic scrubber), but a minority of rows retain some page-furniture residue, and extraction can damage formatting (lost lists, table→pipe residue). Detectors trained on this corpus should compare against AI text formatted the same way (plain extracted text, no HTML).
• Selection bias: English only (lid.176 ≥ 0.65), quality floors as above, and topic × format rebalancing — the corpus is not a natural sample of the 2021 web. Cells that are near-empty on the real web (e.g. Adult × Academic Writing) remain small; see the cell table above and reports/selection_report.json.
• Residual contamination: pre-ChatGPT does not mean zero machine text — 2021 web pages can contain template-generated, machine-translated, or spun/SEO text. The quality floors and junk fences remove most of it, but no guarantee is absolute.
• Content warning: despite URL blocklists and quality filtering, web text may contain offensive, inaccurate, or otherwise objectionable content, and may mention personal data present on public 2021 web pages.
• PII is deliberately NOT masked. This corpus exists to train detectors that distinguish human from AI text; masking emails/IPs with stock replacement strings (as FineWeb does) would inject an "anonymization artifact" that appears only in the human class — a spurious, learnable shortcut — and would make the text something no human actually wrote. The text is redistributed exactly as it appeared on the public 2021 web (as in CommonCrawl itself). Per-row n_emails and n_ips columns count FineWeb-definition PII hits (email regex; public IPv4) so downstream users can filter or apply their own masking if their use case requires it.
• Takedown: to report content (including personal data) for removal, open a discussion on this repo.

License

The curation code and this compilation are released under ODC-By 1.0 (like FineWeb); use of the underlying content remains subject to the CommonCrawl Terms of Use.

Citations

This dataset composes methods from the following works — please cite them when using it:

@inproceedings{penedo2024fineweb,
  title  = {The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale},
  author = {Penedo, Guilherme and Kydl{\'i}{\v{c}}ek, Hynek and Ben allal, Loubna and Lozhkov, Anton and Mitchell, Margaret and Raffel, Colin and Von Werra, Leandro and Wolf, Thomas},
  booktitle = {NeurIPS Datasets and Benchmarks},
  year   = {2024},
  url    = {https://arxiv.org/abs/2406.17557}
}

@article{wang2025ultrafineweb,
  title  = {Ultra-FineWeb: Efficient Data Filtering and Verification for High-Quality LLM Training Data},
  author = {Wang, Yudong and Sun, Zixuan and others},
  journal = {arXiv preprint arXiv:2505.05427},
  year   = {2025},
  url    = {https://arxiv.org/abs/2505.05427}
}

@article{gneissweb2025,
  title  = {GneissWeb: Preparing High Quality Data for LLMs at Scale},
  author = {Gohari, Hajar Emami and Kadhe, Swanand Ravindra and others},
  journal = {arXiv preprint arXiv:2502.14907},
  year   = {2025},
  url    = {https://arxiv.org/abs/2502.14907}
}

@inproceedings{wettig2025weborganizer,
  title  = {Organize the Web: Constructing Domains Enhances Pre-Training Data Curation},
  author = {Wettig, Alexander and Lo, Kyle and Min, Sewon and Hajishirzi, Hannaneh and Chen, Danqi and Soldaini, Luca},
  booktitle = {ICML},
  year   = {2025},
  url    = {https://arxiv.org/abs/2502.10341}
}

Built with datatrove. Data from CommonCrawl. Published 2026-08-31.