Ugiat/madlad-400-es-annotations

MADLAD-400 Spanish annotations: Spain vs Latin America, languages and websites

Annotations for every document of the Spanish split of MADLAD-400 (data-v1p5/es, clean): 250,906,994 documents from 4,993,810 websites. For each document: where it points to in MADLAD-400, the domain group of its website, a Spain vs Latin America score, dialect-marker counts, which languages its lines are in, its website's average score, and whether it passes the Castilian selection rule used by Ugiat Technologies. The text itself is not included: it stays in MADLAD-400, and every row points to it.

• documents: one row per document (250.9M rows); 70,133,412 of them (28.0 %) are selected as Castilian (castilian_selected, see the Castilian selection rule).
• sites: one row per website (4,993,810 rows) with document counts and average scores.
• other_language_lines: 9,929,267 lines of Spanish documents that are confidently in another language (English, Portuguese, French, Catalan…), with their text.

This dataset was prepared by Ugiat Technologies. It is an independent work and is not produced, reviewed or endorsed by Google, AllenAI or the MADLAD-400 authors.

Why this dataset

MADLAD-400, like other web corpora, has one "Spanish" that mixes Spain and Latin America, which together are about a quarter of a billion documents. Anyone who needs one variety (to fine-tune a translation model into Castilian, to build a Latin American language model, to study variation) has to classify it first. This dataset does that once for the whole split and publishes the result, so others can select documents with a filter instead of a multi-day pipeline, and with their own thresholds.

Example

host	tld_group	p_es_ES	site_p_es_ES	markers_vosotros	markers_voseo	castilian_selected
www.autingo.es	spain	1.00	1.00	0	0	true
iresiduo.com	generic	1.00	0.89	0	0	true
www.calarca.net	generic	0.05	0.15	0	0	false
koe.cl	latam	0.05	0.18	0	0	false

Dataset structure

documents

Field	Type	Description
file, line	string, int	The document in MADLAD-400: file under data-v1p5/es/ and 0-based line, revision 9d886a76bd8fa69b294f2dd3843dacb8388ee5a5
url, host, tld	string	From MADLAD-400's url field
tld_group	string	spain (.es, .cat, .eus, .gal, .madrid, .barcelona), near_spain (.eu, .ad), latam (Latin American country domains, including .co), generic (.com, .org, .net, .info… and two-letter domains used generically such as .tv, .io), other_country, unknown
timestamp	string	Crawl timestamp from MADLAD-400
text_sha1_64	string	First 16 hex characters of the SHA-1 of MADLAD-400's text field, to check the pointer
n_chars, n_lines	int	Size of the document
lid_chars	int	Characters in lines of 5+ words, the ones language-identified
es_share	float	Share of lid_chars identified as Spanish (null if no line was identified)
langs	list of {lang, chars}	Characters per language identified by fastText lid.176 (und: probability < 0.5)
other_langs	list of strings	Languages other than Spanish with at least 5 % of lid_chars
norm_tokens	int	Tokens seen by the variety classifier after normalisation
p_es_ES	float	Probability that the document is Spain's Spanish, from Ugiat/spanish-variety-fasttext, computed on its Spanish lines; null when norm_tokens < 20
markers_vosotros	int	Matches of vosotros / vosotras / sois / os / verb forms in -áis, -éis
markers_voseo	int	Matches of vos, sos, tenés, querés, podés, sabés, hacés, mirá, fijate, andá
markers_es_vocab	int	Matches of ordenador, zumo, coche
markers_latam_vocab	int	Matches of computadora, celular, jugo, ahorita, platic…
site_docs, site_p_es_ES	int, float	Documents of the website in this split and their mean p_es_ES (from sites)
castilian_selected	bool	true if the document is selected as Castilian (Spain's Spanish): it comes from a Spanish domain or from a generic website that scores as Spanish, its own p_es_ES ≥ 0.9, and it has no voseo or Latin American vocabulary. The rule favours precision: false does not mean Latin American, many Castilian documents (neutral text, small websites) are simply not selected. Details in the Castilian selection rule

sites

host, tld_group, docs, docs_scored (with a score), site_p_es_ES (mean score), docs_p_ge_0_9, chars.

other_language_lines

file, line, line_idx (0-based line within the document, splitting MADLAD-400's text on its \n separators), lang, lid_prob (≥ 0.9), text.

Statistics

tld_group	Documents	Text	Score ≥ 0.9	Score < 0.1
generic	162,374,104	724.5 GB	38.5 %	28.0 %
latam	44,618,146	153.7 GB	2.4 %	79.5 %
spain	37,878,024	153.3 GB	75.5 %	2.3 %
other_country	4,505,737	24.1 GB	30.8 %	18.4 %
near_spain	1,529,360	7.5 GB	63.3 %	3.5 %
unknown	1,623	< 0.1 GB
total	250,906,994	1,063 GB

Languages of the identified lines (by characters): Spanish 98.1 %, undetermined 1.5 %, English 0.22 %, Portuguese 0.03 %, French 0.02 %, Catalan 0.02 %, Italian 0.01 %, German 0.01 %.

Generic-domain websites with ≥ 20 documents and site_p_es_ES ≥ 0.85: 241,801 sites, 51.3M documents (214.5 GB), e.g. elespanol.com, lavanguardia.com, antena3.com, elconfidencial.com.

Loading

from datasets import load_dataset

docs = load_dataset("Ugiat/madlad-400-es-annotations", "documents", split="train", streaming=True)
castilian = docs.filter(lambda r: r["castilian_selected"])
latam = docs.filter(lambda r: r["tld_group"] == "latam" and (r["p_es_ES"] or 1) < 0.1)

Use file and line to fetch the text from MADLAD-400 (download the shard, read line line, check text_sha1_64).

How the annotations were produced

• Domain groups from the URL's top-level domain.
• Language identification: fastText lid.176 on every line of 5+ words (4.7 billion lines).
• Variety score (p_es_ES): a fastText classifier trained on 912,477 documents per class from 200 random shards, labelled by domain (.es vs Latin American country domains, sites held out for validation). Before scoring, text is lowercased and capitalized words, digits and URLs are removed, so that the classifier learns how the text is written rather than which places or brands it mentions. On held-out websites it reaches precision 0.96 / recall 0.74 for Spain at a 0.9 threshold (against domain labels). Details: Ugiat/spanish-variety-fasttext.
• Markers: regular expressions, counted on the whole document; they are not used by the classifier and serve as an independent check: documents scored ≥ 0.9 have 175 vosotros-marker documents per 1,000 and 5 voseo per 1,000, against 12 and 32 for documents scored < 0.1.

The Castilian selection rule

castilian_selected is true when all of these hold:

1. Where it comes from: the website is on a Spanish domain (tld_group spain or near_spain: .es, .cat, .eus, .gal, .madrid, .barcelona, .eu, .ad), or it is a generic website (.com, .org, .net…) with site_docs ≥ 20 and site_p_es_ES ≥ 0.85 — 241,801 such websites, e.g. elespanol.com, lavanguardia.com, elconfidencial.com.
2. The document itself scores as Spain's Spanish: p_es_ES ≥ 0.9.
3. No Latin American marks: markers_voseo = 0 and markers_latam_vocab = 0.

It selects 70,133,412 documents (28.0 % of the split). Each condition alone lets Latin American text through (all .es documents ≈ 5 %, p_es_ES ≥ 0.9 on any domain ≈ 4.5 %); combined, none of the 72 selected documents in an internal reference set of 340 annotated documents was Latin American (≲ 4 % at 95 % confidence).

It is a high-precision selection for building Castilian training data, not a classification of every document: false means "not selected", not "Latin American". To select Latin American Spanish or to trade precision for coverage, combine tld_group, p_es_ES, site_p_es_ES and the markers_* columns with your own thresholds. This is the rule that produced Ugiat/madlad-400-es_ES-sentences.

Uses

• Selecting Spain's Spanish (or Latin American Spanish) from MADLAD-400 for training or evaluation.
• Finding multilingual pages (e.g. Spanish–Catalan, Spanish–Basque institutional sites) for parallel-data mining, with other_langs and other_language_lines.
• Studying the composition of a large Spanish web corpus by website, domain and variety.

Limitations

• The labels are automatic. Domain labels are noisy (Latin American authors on Spanish sites and vice versa); the classifier learned from them and is only reliable per document, not per sentence.
• Neutral text has no variety. Many documents show no dialect evidence; their score reflects weak cues and should not be read as certainty.
• generic websites are judged by their average score, which the same classifier produces.
• Language identification mislabels some short or name-heavy lines (e.g. Spanish lines with Catalan place names as Catalan) and confuses Galician with Spanish and Portuguese.
• other_language_lines contains text from web pages, not scrubbed of personal data.

Licensing, attribution and terms of use

• License: ODC-BY 1.0, like MADLAD-400 (whose README also states CC BY 4.0). Give appropriate credit to MADLAD-400 and to this dataset.
• Common Crawl. Use is also subject to the Common Crawl terms of use.
• Takedown. Write to the contact below to have content removed.

Please cite the original corpus

@misc{kudugunta2023madlad400,
      title={MADLAD-400: A Multilingual And Document-Level Large Audited Dataset},
      author={Sneha Kudugunta and Isaac Caswell and Biao Zhang and Xavier Garcia and Christopher A. Choquette-Choo and Katherine Lee and Derrick Xin and Aditya Kusupati and Romi Stella and Ankur Bapna and Orhan Firat},
      year={2023},
      eprint={2309.04662},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Suggested reference for this dataset:

@misc{madlad400_es_annotations,
      title={MADLAD-400 Spanish annotations: Spain vs Latin America, languages and websites},
      author={Ugiat},
      organization={Ugiat Technologies},
      year={2026},
      howpublished={\url{https://huggingface.co/datasets/Ugiat/madlad-400-es-annotations}},
      note={Derived from MADLAD-400, ODC-BY 1.0}
}

Contact

Ugiat Technologies, ugiat@ugiat.com.