baobabtech/millid-dataset

MillID: language identification for 1,000+ languages

MillID (MultiLingual Language IDentifier) is the dataset behind the MillID language identifiers (MillID tiny v1.5, MillID medium v1, MillID large v1.5; compared with GlotLID and fastText on the leaderboard). Passages of text labelled with their language, for training and evaluating a language identifier: 1,641 language-script labels (1,551 ISO 639-3 languages, 56 scripts), 2,390,082 passages, plus the Universal Declaration of Human Rights in 429 languages as a human-labelled test set.

The MillID models were trained and scored on a fuller build that also holds Save the Children's (23,104 passages) and Nal'ibali's (16,500 passages) text, which is not ours to redistribute. This dataset is that build without it: every other passage keeps its split and every label its id, so test here is the models' test split without those passages.

Built by Baobab Tech to identify the language of real documents, so most of the text comes from PDFs, cut the way a page is read: short lines, sentences and paragraphs. Code: baobab-tech/millid.

Configs

Config	Split	Rows	What it is
default	train	1,911,469
default	validation	241,469
default	test	237,144	split by document from train, never by passage
udhr	test	27,623	UDHR-LID, labelled by the translators; 358 of its 429 labels are in default

from datasets import load_dataset

ds = load_dataset("baobabtech/millid-dataset")            # train / validation / test
udhr = load_dataset("baobabtech/millid-dataset", "udhr")  # human-labelled test set

Fields

Field
text	the passage
label	ISO 639-3 + ISO 15924 script, e.g. eng_Latn, amh_Ethi, srp_Cyrl
language, iso639_3, iso639_1, script	the label spelled out
size	short (20 to 80 characters), sentence (up to 300) or paragraph (up to 1,000)
source, doc_id, page, url, title	where it came from
license, attribution	per row: filter to what your use allows, and credit accordingly
label_quality	publisher (the publisher's label), verified (GlotLID agrees), declared (see below)
glotlid, glotlid_prob	GlotLID's reading of the passage

Sources

Source	What	Passages	Labels from	Licence
HuggingFaceFW/finepdfs	PDF text from Common Crawl	605,480	GlotLID, per page	ODC-By 1.0
HuggingFaceFW/fineweb-2	Web text from Common Crawl	1,648,616	GlotLID	ODC-By 1.0
cis-lmu/GlotStoryBook	Children's storybooks: African Storybook, Pratham Books, LIDA Stories, Little Cree Books	91,800	the publishers	per story: CC BY, CC BY-NC, CC BY-NC-SA, public domain
cis-lmu/GlotSparse	News sites in sparse languages	15,000	the sites	CC0 1.0
CAWST resources	Water, sanitation and hygiene training material	29,186	CAWST	CC BY / CC BY-SA (CAWST open content)

How it was made

• Passages. Text is NFKC-normalised, which folds the ligatures and Arabic presentation forms PDF text layers are full of; URLs, emails and dot leaders are removed; a passage must be mostly letters. Each page gives passages of three sizes, because short text is where language identification is hardest. At most 40 passages come from one PDF, 20 from one article and 10 from one web page. Exact duplicates are removed across every source.
• Labels. Publisher-labelled sources keep their label unless GlotLID is confident (≥ 0.8) a passage is in another major language, typically a copyright line or a credit (publisher). FinePDFs and FineWeb-2 were labelled by GlotLID per document, so a passage from them is kept only when GlotLID's reading of the passage agrees (verified); for a language whose data GlotLID agrees with less than 70% of the time, the label being in its top three is enough (declared).
• Text that is no language is removed, whatever its label: passages GlotLID is confident (≥ 0.8) are zxx (no linguistic content). In these sources that is mostly a PDF whose font maps letters to the wrong code points, such as a Russian document whose text layer reads ɉɪɟɫɥɟɞɨɜɚɧɢɟ for Преследование. So is a document catalogued in the wrong language (a Yemen report catalogued as Estonian).
• Coverage. A label needs at least 100 passages from two documents; 310 labels had less and are not included. Each source contributes at most a fixed number of passages per label, so the large languages do not dominate.
• Splits are by document, so no test passage shares a document with a training one. A label with fewer than five documents is split by page.
• UDHR is kept out of training. The Declaration is published everywhere and the crawls hold copies of it. 4,391 passages sharing a 30-character run of letters with a UDHR-LID sentence were removed, so udhr tests recognition, not memory.

The smallest labels:

Label	Language	Train	Validation	Test	Documents
aeb_Latn	Tunisian Arabic	61	38	1	15
plw_Latn	Brooke's Point Palawano	77	18	8	14
sba_Latn	Ngambay	84	20	0	13
lis_Lisu	Lisu	92	7	6	15
ade_Latn	Adele	31	43	31	6
old_Latn	Mochi	82	13	11	31
nre_Latn	Southern Rengma Naga	54	10	42	8
rki_Mymr	Rakhine	89	20	0	13

labels.csv lists every label with its counts per split and source.

Limitations

• Labels from FinePDFs and FineWeb-2 are GlotLID's. A model trained on them learns GlotLID's labels, mistakes included, so udhr is the fair test. GlotLID was itself trained on UDHR and the Glot corpora, so its own score on them is optimistic.
• Genre. For many low-resource languages the web is mostly religious text, and the storybooks are children's books. A model can learn the genre along with the language.
• Close languages (Bemba and Lamba, Zulu and Ndebele, Arabic varieties) are hard by nature; short passages of them are often genuinely ambiguous.

Licensing

There is no single licence. Every row carries its own in license, and its credit in attribution:

• ODC-By 1.0: 2,254,096
• CC BY: 81,221
• CAWST open content (resources.cawst.org/cc): 20,680
• CC0 1.0: 15,000
• CC BY-NC: 6,277
• CC BY-SA 4.0: 4,602
• CC BY-NC-SA: 4,146
• CC BY 3.0: 1,949
• CC BY-SA: 1,787
• CC BY 4.0: 168
• Public Domain: 156

FinePDFs and FineWeb-2 are ODC-By 1.0, also subject to Common Crawl's terms of use. UDHR-LID is CC0. Rows marked NC are for non-commercial use only.

Citation

Please cite the sources: FinePDFs and FineWeb-2 (Hugging Face), GlotLID and GlotStoryBook / GlotSparse (Kargaran et al., CIS LMU Munich), and UDHR-LID.