UDHR

The Universal Declaration of Human Rights in every translation of the UDHR collection formerly at unicode.org/udhr, split into units that are parallel across languages, with English as the base.

• 532 translations: 462 language_script codes, 446 ISO 639-3 languages, 43 scripts
• 47,980 units: every translation has all 30 articles (521 in the source; for the others the absent articles are empty rows, see Missing text)
• every translation file of the collection is in, including partial ones and several translations of one language, except one exact duplicate: the Latin-script Central Kurdish file (ckb) has the same text as the Northern Kurdish one (kmr), byte for byte, and is left out

Source: the collection's XML files (formerly at unicode.org/udhr), version of 2026-07-28. The 82 entries of its index without a translation file (stage 2) have no text and are not included. Built with build_udhr.py (in this repository).

How this differs from cis-lmu/udhr-lid: udhr-lid is a language-identification test set (cleaned paragraphs, a few files removed, not in order, not parallel). This dataset keeps every translation and its structure, so it is parallel and complete.

Subsets

One subset per translation, named after its language code (lang) and, where a language has several translations, the variant part of the collection's id: deu_Latn_1901 and deu_Latn_1996, cmn_Hans, cmn_Hans_beijing, cri_Latn_007 (numeric ids), und_Latn_013. A language with one translation is just its code, e.g. fra_Latn. The collection's id is in the key column. all (the default) has every translation and translations one row per translation.

from datasets import load_dataset
de = load_dataset("cis-lmu/udhr", "deu_Latn_1996", split="test")
everything = load_dataset("cis-lmu/udhr", split="test")   # subset "all"

Units

The same structural id in every translation:

unit	content
title	title of the declaration
preamble.title	heading of the preamble
preamble.recital.1 ... preamble.recital.7	the seven recitals ("Whereas ...")
preamble.proclamation	the rest of the preamble ("Now, therefore, The General Assembly Proclaims ..."), its paragraphs joined with a space
preamble.p1, preamble.p2, ...	preamble paragraphs, in translations where the recitals cannot be told apart (then aligned is false)
article.K.title	heading of article K (1-30)
article.K.pM	paragraph M of article K
article.K.iM	item M of the numbered list of article K (more paragraphs in one item: article.K.iM.pJ)
note.N.pM	paragraph M of the N-th translator's note (no English counterpart)

English has 90 units. The recitals are recognised when the first seven of eight or more preamble paragraphs share an opening word that the eighth does not ("Whereas", "Considérant", "Da", "鉴于", ...) or have the lengths of the English recitals (scaled by how long the translation's articles are): 331 of the 505 translations with a preamble. Translations split the proclamation into one to six paragraphs, so it is one unit.

Columns

all and the per-translation subsets (one row per unit and translation, in document order):

column
subset	the subset of the translation (see Subsets)
unit	structural id, as above
lang	ISO 639-3 code and ISO 15924 script, e.g. deu_Latn, cmn_Hans, uig_Arab; corrected where the collection's code is wrong (see Corrections)
legacy_lang	the code as in the collection (equal to lang except for the corrected translations)
relabel	why lang differs from legacy_lang (empty otherwise)
key	the collection's translation key (file udhr_<key>.xml); distinguishes several translations of one language, e.g. deu_1901 and deu_1996
name	language name in the collection
iso639_3	the language of lang (und where the collection does not identify the language)
iso15924, bcp47, dir	as in the collection's index
stage	the collection's completeness stage: 4 = XML version, reasonably complete; 3/2 = source document only partly in XML; 1 = no source yet
ohchr	OHCHR page or id of the translation, where known
text	the text (whitespace normalised); empty when missing
missing	there is no text for this unit in this translation (see Missing text)
issue	why: placeholder, english or absent (empty when there is text)
english	the English text of the same unit
aligned	the unit has an English counterpart and its section (the preamble, or one article) has the same paragraphs and list items as in English. When a translation merges or splits paragraphs (most often the two paragraphs of article 2), the units of that section are kept but marked false. 94% of all units are aligned (missing units are not); the median translation is 98.9% aligned

translations (one row per translation, 532 rows): subset, key, lang, legacy_lang, relabel, name, iso639_3, iso15924, bcp47, dir, stage, ohchr, units, missing (units without text), aligned (share of its units), recitals (the preamble is split into recitals), articles (number of articles with text: 521 translations have all 30).

Missing text

A unit has no text (missing true, text empty) in three cases, given in issue:

issue	the source has	units	translations
placeholder	a placeholder instead of the text: [missing], [Missing?], (missing), ?, @, an empty paragraph, or only an editor's note in square brackets ([Missing 15.2], [Art. 29.3 is not available], [Ban Ki-moon's signature])	138	61
english	English in place of the translation: the Ao Naga file's notes (njo, note.1.p2-p7, an English leaflet before its Ao version). Found by running GlotLID over every unit and reading the hits; these are the only ones	6	1
absent	the whole article is not in the file (e.g. Chuvash has articles 1-19 only): one empty row per English unit of the article, in English order, so every translation has all 30 articles	425	11

Units that are only missing because a translation merged paragraphs (the text is in the neighbouring unit) are not empty rows: those sections have aligned false. Headings identical to English ("Article 1") are kept: they are right in French, Catalan and Occitan.

Corrections

Four translations have a language code in the collection that their text shows to be wrong. They are kept, with the corrected code in lang, the collection's code in legacy_lang and the reasons in relabel (the same changes as in cis-lmu/udhr-lid). Each change rests on an outside source and on the text itself:

key	legacy_lang	lang	outside source	in the text
nbl	nbl_Latn	nde_Latn (North Ndebele)	OHCHR's page for this text (LangID nel) gives the source as the United Nations Information Centre, Zimbabwe; its library links the recording as udhr.audio/...?lng=nde. The collection names it only "Ndebele"	North Ndebele njalo and kumbe (about 100 per 10,000 words); none of the South Ndebele begodu, ngombana, kodwana, namkha
krl	krl_Latn	olo_Latn (Livvi)	OHCHR lists it (LangID krl, Ristikanzan oigevuksien yhtehine deklaratsii, 2009) only as "Karelian" and notes there is no standard Karelian, each writer using his own dialect. English Wikipedia's Karelian article quotes its article 1 as the sample of Olonets (Livvi) Karelian	Livvi words (oigevus, ristikanzu), -zen forms; š at 4 per 1,000 words against about 250 in Karelian Proper
pcd	pcd_Latn	wln_Latn (Walloon)	OHCHR itself names this text "Picard" (LangID frn2, from Diffusion Multilingue des Droits de l'Homme, Belgium), and the collection follows it; but its spelling (å, dj) is that of OHCHR's Walloon translation (frn1), and its words are in the DTW Walloon dictionary (todi, cisse)	Walloon cisse, djins, todi, k'dûhance, cwèri; none of the Picard ech, ed, chés. A second Walloon translation next to wln
tzm	tzm_Latn	kab_Latn (Kabyle)	OHCHR's page for this text (LangID tzm) gives the source as the Observatoire National des Droits de l'Homme, Algeria (Central Atlas Tamazight is spoken in Morocco)	Article 1 is almost word for word the collection's Kabyle translation (071): Imdanen, akken ma llan ttlalen d ilelliyen ... d yizerfan

Left out: the Latin-script Central Kurdish file (ckb) has the same text as the Northern Kurdish one (kmr), byte for byte.

Usage

from datasets import load_dataset
udhr = load_dataset("cis-lmu/udhr", split="test")

# German (1996) next to English, aligned units only
de = load_dataset("cis-lmu/udhr", "deu_Latn_1996", split="test").filter(lambda r: r["aligned"])
print(de[20]["unit"], de[20]["english"], de[20]["text"], sep="\n")

# a parallel table: one column per translation
import pandas as pd
df = udhr.to_pandas()
wide = df[df.aligned].pivot(index="unit", columns="subset", values="text")

Use as a test set

UDHR is one of the most translated documents, and parts of it are on Wikipedia, Tatoeba and many web pages. Use it for evaluation, not training, and remove any sentence of it from your training data.

License

The UDHR is copyright-free. We license the packaging, the unit ids and the metadata under CC0 1.0. The texts come from the UDHR collection formerly at unicode.org/udhr.

Citation

We maintain this dataset for GlotLID: GlotLID uses it, and it is the upstream of cis-lmu/udhr-lid, the UDHR test set of GlotLID. The language corrections above come from that work. If you use this dataset, please cite GlotLID:

@inproceedings{
  kargaran2023glotlid,
  title={{GlotLID}: Language Identification for Low-Resource Languages},
  author={Kargaran, Amir Hossein and Imani, Ayyoob and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
  booktitle={The 2023 Conference on Empirical Methods in Natural Language Processing},
  year={2023},
  url={https://openreview.net/forum?id=dl4e3EBz5j}
}