Yusser/CC-LARD-topics

CC-LARD — Common Crawl Language and Region Dataset

with Cultural Topic Assignments and CTO leaf alignment.

Companion corpus for the paper "CC-LARD: A Common Crawl Language and Region Dataset with Cultural Topic Assignments and a 175-Benchmark Survey" (under review, NeurIPS Datasets & Benchmarks 2026).

⚠️ License at a glance

This dataset is released under ODC-By v1.0 plus the Common Crawl Foundation Terms of Use, following the licensing pattern of FineWeb and Dolma.

✅ You may	❌ You must
Train models on it (subject to the Common Crawl ToU indemnification clause for AI / ML / LLM use)	Cite the CC-LARD paper and Common Crawl
Redistribute, modify, and build derivatives	Pass through this license to derivatives
Use it for academic, commercial, and applied purposes	Independently comply with original content creators' rights and any embedded third-party terms

Note: the companion benchmark-survey artefact at Yusser/survey-culture-topics is under a more restrictive research-only license that prohibits use as training data. The two repos are released under different terms; do not conflate them.

Dataset summary

This HuggingFace repository releases the topic-assigned subset of CC-LARD: 1,239,169 documents drawn from the 53 topic-modelled locales, each carrying its locale label, language, FASTopic topic, and CTO leaf assignment. The full CC-LARD corpus (13.15 M documents over 4,334 locales) is regenerable from the released pipeline code and Common Crawl snapshot list.

Statistic (full CC-LARD)	Value
Documents (full corpus)	13,150,911
Languages	140
Locales	4,334
Regions	349
Topic-modelled locales	53
Documents in this HF release (topic-assigned subset)	1,239,169
FASTopic topics across 53 locales	1,634
CTO leaves used as the shared index	14
Common Crawl snapshots	41
Pipeline retention (28.1 M → 13.15 M)	46.74 %

Companion repo

	Repo	License
Corpus side (this repo)	Yusser/CC-LARD-topics	ODC-By + CC ToU
Survey side	Yusser/survey-culture-topics	Research-only, no training (gated access)

Schema (data/<locale>/train.parquet)

Column	Type	Description
text	string	Document text (post-extraction, post-quality, post-dedup)
lang	string	Detected language code (GlotLID, ISO-639-1/3)
region	string	Resolved region (ISO-3166-1 alpha-2, or null if XX)
locale	string	lang-REGION code (e.g., en-IN); XX if unresolved
url	string	Source URL
lang_confidence	float64	GlotLID detection confidence (≥ 0.65 by filter threshold)
region_source	string	How the region was resolved (HTML lang tag / HTTP header / ccTLD / etc.)
warc_date	string	WARC-record date (snapshot timestamp)
text_length	int64	Character length
title	string	Document title (first heading or <title>)
text_hash	string	SHA-256 of the post-extraction text (used for exact dedup)
dominant_topic_id	int64	FASTopic topic id within the locale
dominant_topic_label	string	FASTopic top-words label
dominant_topic_description	string	LLM-generated topic description
dominant_topic_cto_category	string	One of the 14 CTO leaves (e.g., L2_artifacts, L3_values_bias)
dominant_topic_probability	float64	Probability mass on the dominant topic
topic_distribution	list[float64]	Full FASTopic topic-distribution vector for the document

Construction pipeline (summary)

Common Crawl WARC →
  Trafilatura HTML extraction (per shard, 128-core parallel) →
    GlotLID v3 language ID (confidence ≥ 0.65) →
      Adult-domain filter →
        Quality filter (length, boilerplate, alpha-ratio) →
          Exact dedup (SHA-256) → Near-dedup (MinHash-LSH, Jaccard ≥ 0.8) →
            Length filter (≥ 200 chars) → Token-count filter (≥ 20 tokens)
            ↓
       Per-locale text shards (the CC-LARD core)
            ↓
       FASTopic per locale (K = 1 topic per ~100 docs, capped 30..500)
            ↓
       GPT-4o-adjudicated topic → CTO leaf mapping
            ↓
       This dataset: per-document CTO leaf labels for 53 high-resource locales

Full retention table:

Stage	Input	Output	Drop
02 LangID	28,136,673	22,101,198	1,781,943
03 Adult filter	22,101,198	20,113,722	205,533
05 Quality filter	20,113,722	17,800,447	2,313,275
06 Exact dedup	17,800,447	15,276,796	2,523,651
07 Near dedup	15,276,796	14,814,217	462,579
08 Length filter	14,814,217	13,492,792	1,321,425
09 Token-count filter	13,492,792	13,150,911	341,881

Recommended uses

• Locale-aware corpus auditing: which cultural topics are over- / under-represented in which locales.
• Building locale-level baselines for cultural evaluation.
• Pretraining-data analysis: comparing CC-LARD CTFs against your own pretraining corpus to identify cultural blind spots before training.
• Cross-locale fairness analysis: diagnosing which locales are present in the training distribution but missing from evaluation.

Out-of-scope uses

• Treating CC-LARD as a complete or unbiased sample of cultural content in any locale: it is filtered, deduplicated, and limited to documents that expose locale metadata (47 % of full CC-LARD is XX-locale and is excluded from topic modelling and from this HF release).
• Treating per-document CTO-leaf labels as ground truth: 31 % of GPT-4o topic assignments override the embedding prior, and 8 % of sampled assignments are flagged "incorrect" by human spot-check (see paper).
• Using locale labels as a proxy for speaker community: en-IN indicates Indian-English web content (institutional, business, news), not Indian English speakers.

Privacy and takedown

Despite filters for common PII patterns (emails, public IPs, phone-number-shaped strings), residual personal data may remain. To request removal of personal data or any document whose URL matches your content, contact the CC-LARD authors (see paper). We will remove the document on receipt of a verifiable request and publish a versioned release log of removals.

Citation

@inproceedings{cclard2026,
  title     = {{CC-LARD}: A {Common Crawl} Language and Region Dataset
               with Cultural Topic Assignments and a 175-Benchmark Survey},
  author    = {Anonymous},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  note      = {Datasets and Benchmarks track, under review}
}

Please also cite the Common Crawl Foundation as the source of the underlying web text:

Common Crawl Foundation. https://commoncrawl.org

Acknowledgements

Built on the Common Crawl Foundation's WARC archives and on open-source tooling: Trafilatura, GlotLID, FASTopic, BGE-M3, BERTopic, MinHash-LSH. See the paper appendix for full version pinning and compute details.