Hugging Face's logo

Datasets:
allenai
/
MADLAD-400

Tasks:
Text Generation
Size:
n>1T
ArXiv:
License:
Dataset card Files Community
10
MADLAD-400 / README.md
drschwenk's picture
drschwenk
Update README.md
e2d81d9
preview code
|
raw
history blame
143 kB
metadata
license: odc-by
task_categories:
  - text-generation
size_categories:
  - n>1T

MADLAD-400

Dataset and Introduction

MADLAD-400 (Multilingual Audited Dataset: Low-resource And Document-level) is a document-level multilingual dataset based on Common Crawl, covering 419 languages in total. This uses all snapshots of CommonCrawl available as of August 1, 2022. The primary advantage of this dataset over similar datasets is that it is more multilingual (419 languages), it is audited and more highly filtered, and it is document-level. The main disadvantage is also its strength -- being more filtered, it may lack the recall needed for some applications.

There are two versions released: the noisy dataset, which has no filtering except document-level LangID, and the clean dataset, which has a variety of filters applied, though it naturally has a fair amount of noise itself. Each dataset is released in a document-level form that has been deduplicated.

Loading

You can load both the clean and noisy versions of any language by specifing its LangID:

madlad_abt = load_dataset("allenai/madlad-400", "abt")

A list of langagues can also be supplied with a keyword argument:

madlad_multilang = load_dataset("allenai/madlad-400", languages=["abt", "ace"])

Additionally, you can load the noisy and clean subsets seperately with the split keyword argument:

madlad_multilang_clean = load_dataset("allenai/madlad-400", languages=["abt", "ace"], split="clean")

LangID model and Crawl

Following Language Id In the Wild, we trained a Semi-Supervised LangId model (SSLID) on 500 languages. The training data is as described in that paper, with the differences that 1) training data is sampled to a temperature of T=3 to reduce over-triggering on low-resource languages; and 2) the data is supplemented with web-crawled data from the same paper (that has already been through the various filters described therein) in the hopes that it will increase robustness to web-domain text.

Filtering

Before separating the raw CommonCrawl corpus by LangID, these filtering steps are done, similar to Raffel et al (2020):

  • Discarded any page with fewer than 5 sentences and only retained lines that contained at least 3 words.
  • Removed any line with the word Javascript.
  • Removed any page where the phrase “lorem ipsum” appeared.
  • Removed any pages containing the phrases "terms of use", "privacy policy", "cookie policy", "uses cookies", "use of cookies", "use cookies"
  • Removed any pages that contained a curly bracket.
  • To deduplicate the data set, discarded all but one of any three-sentence span occurring more than once in the data set.

The noisy subset of the data was filtered only by document-level LangID, which was taken to be the majority sentence-level LangID prediction. The clean subset removed all documents with a percent_questionable score greater than 20%. It furthermore removed any document with under 5 sentences.

The pct_questionable score is simple the percentage of sentences in the input document that were "questionable". A sentence was considered questionable if any of the following were true:

  • LangID Consistency: the sentence-level LangID does not match the document-level LangID
  • List Case: The sentence has at least 12 tokens, and over 50% percent of the tokens began in a capital letter.
  • Length: The sentence has under 20 characters or over 500 characters (note: this is a bad heuristic for ideographic languages)
  • Danger Chars: Over 20% of the characters in the sentence match [0-9{}+/()>]
  • Cursedness: The sentence matches a cursed regex (see below)

Cursed Substrings

Based on the initial round of data audits, the authors created a heuristic list of substrings and regexes accounting for a large amount of questionable content. Keep in mind that these all are fed into the pct_questionable score -- a sentence is only excluded from the clean dataset if over 20% of the sentences in that document are flagged as questionable.

notes about cursed substrings:

  • low quality sentences ending in the pipe character were very common. Before you ask, this was not Devanagari-script text using a Danda.
  • The last few regexes are meant to match A N T S P E A K, List Case, and weirdly regular text (for instance, lists of shipping labels or country codes)
# this implementation is for demonstration and is pretty inefficient;
# to speed it up, use string inclusion (`in`) instead of regex for all but the
# last four, and for those use a compiled regex.
def is_cursed(s):
  return any(re.findall(curse, s) in s for curse in CURSED_SUBSTRINGS)

CURSED_SUBSTRINGS = [" №", "���", "\\|\\s*$", " nr\\.$", "aute irure dolor ", " sunt in culpa qui ", "orem ipsum ", " quis nostrud ", " adipisicing ", " dolore eu ", " cupidatat ", "autem vel eum", "wisi enim ad", " sex ", " porn ", "黄色电影", "mp3", "ownload", "Vol\\.", " Ep\\.", "Episode", " г\\.\\s*$", " кг\\.\\s*$", " шт\\.", "Develop", "Facebook", " crusher ", " xxx ", " ... ... ... ... ... ... ... ... ...", " .... .... .... .... .... .... .... .... ....", " [^ ] [^ ] [^ ] [^ ] [^ ] [^ ] [^ ] [^ ] [^ ]", ", ..,,? ..,,? ..,,? ..,,?"]

Virama Correction

Many languages using Brahmic Abugida (South and Southeast Asian scripts like Devanagari, Khmer, etc.) use some variant on the virama character. For whatever reason, it was found that this character was often messed up in the common crawl snapshots used. Therefore, for the languages bn my pa gu or ta te kn ml si th tl mn lo bo km hi mr ne gom as jv dv bho dz hne ks_Deva mag mni shn yue zh ja kjg mnw ksw rki mtr mwr xnr, a special correction step was done.

For these languages, the authors took the list of all virama characters and removed all unnecessary spaces between each instance of a virama character and the next character with a regex.

'%s' % regex.sub(r' ([%s]) ' % _VIRAMA_CHARS, '\\1', x)

Myanmar Font Compatibility

Prior to 2019, the most popular font for Burmese websites was the Zawgyi font. The authors used Myanmar Tools to convert text.

Several scripts, like the Chinese script, Tibetan script, and Thai, do not use whitespace to separate characters. The languages with this property in this dataset are yue zh ja th lo kjg mnw my shn ksw rki km bo dz.

Alas, the Length aspect of the pct_questionable score was calculated using simplistic whitespace tokenization, and therefore rendered the whole pct_questionable score invalid for those languages. Therefore, for these languages, the "clean" data is identical to the "noisy" data (barring Chinese; see below.)

Special filters

Chinese had a particular issue with pornographic content. After manual inspection a list of strings likely to be present in pornographic content was developed. All pages containing at least one of these strings were removed. Resulted in 17% reduction in number of documents and 56% reduction in file size.

pornsignals = "caoporn caoprom caopron caoporen caoponrn caoponav caopom caoorn 99re dy888 caopro hezyo re99 4438x zooskool xfplay 7tav xxoo xoxo 52av freexx 91chinese anquye cao97 538porm 87fuli 91pron 91porn 26uuu 4438x 182tv kk4444 777me ae86 91av 720lu yy6080 6080yy qqchub paa97 aiai777 yy4480 videossexo 91free 一级特黄大片 偷拍久久国产视频 日本毛片免费视频观看 久久免费热在线精品 高清毛片在线看 日本毛片高清免费视频 一级黄色录像影片 亚洲男人天堂 久久精品视频在线看 自拍区偷拍亚洲视频 亚洲人成视频在线播放 色姑娘综合站 丁香五月啪啪 在线视频成人社区 亚洲人成视频在线播放 久久国产自偷拍 一本道 大香蕉无码 香港经典三级 亚洲成在人线免费视频 天天色综合网 大香蕉伊人久草 欧美一级高清片 天天鲁夜夜啪视频在线 免费黄片视频在线观看 加比勒久久综合 久草热久草在线视频 韩国三级片大全在线观看 青青草在线视频 美国一级毛片 久草在线福利资源 啪啪啪视频在线观看免费 成人福利视频在线观看 婷婷我去也 老司机在线国产 久久成人视频 手机看片福利永久国产 高清国产偷拍在线 大香蕉在线影院 日本高清免费一本视频 男人的天堂东京热 影音先锋男人资源 五月婷婷开心中文字幕 亚洲香蕉视频在线播放 天天啪久久爱视频精品 超碰久久人人摸人人搞".split()

A few more random notes, comparing to common alternative codes for these languages:

  • fil for Filipino/Tagalog, not tl
  • ak for Twi/Akan, rather than tw. This includes Fante.
  • Unfortunately use the macro code chm for Meadow Mari (instead of the correct mhr), and mrj for Hill Mari
  • no for Norwegian Bokmål, whereas some resources use nb
  • ps for Pashto instead of pbt (Southern Pashto)
  • ms for Standard Malay, not zlm
  • sq for Albanian, and don't distinguish dialects like Gheg (aln) and Tosk (als)
  • ber as the code for Tamazight, after consultation with Tamazight speakers opining that the dialect distinctions are not significant. Other resources use the individual codes like tzm and kab.
  • Macrocode qu for Quechua. In practice, this seems usually to be a mix of the Ayacucho and Cusco dialects. Other resources, like NLLB, may use the dialect code, e.g. quy for Ayacucho Chanka. The same is true for a few other macro codes, like ff (Macro code for Fulfulde, whereas other sources may use e.g. fuv.)
  • Really, there are notes that can be made about almost any code, from the well-accepted conventions like zh for Mandarin, to many dialectical notes, like which variant of Hmong really is the hmn data? But the above ones are made specifically for ones where the authors are aware of other datasources floating out there that use different conventions.

Audit

Following Quality at a Glance, the authors performed an "audit" of every corpus in this dataset. Although the authors did not speak most languages, they were able to give high-level comments on the general quality. They looked at a sample of 20 documents of each language.

After an initial round of auditing, they devised a new set of filters and applied them. They then re-did all audits.

Overall notes from the audit

The decision was to include languages that looked noisy, but omit any language that was clearly majority noise, or only had 20 or fewer docs. This is a low bar -- twenty documents can be very little indeed, and some of the corpora released are quite noisy, but all of them should have at least the potential to be used in some useful way. The motivation for not releasing nonsense or tiny datasets is to not give a false sense of how multilingual this dataset actually is ("Representation washing"), as recommended by Quality at a Glance.

A few overarching points:

  • Many low-resource languages only had Bible text, or in some cases jw.org data. These are marked in the rows below. Generally ok bible means that 100% of the audited sentences were Biblical, whereas if bible is simply mentioned in the note, it was not the only source of data.
  • Indian languages in the Latin script had a high concentration of pornographic content.

Renames and Merges as a result of the Audit

In several cases, it was clear from the audit that the corpora were not in the languages that the LangID model claimed they were. This led to the following renames:

  • dty renamed to zxx-xx-dtynoise, aka a "language" of noise. This is mainly mis-rendered PDFs and may have some practical applications for decoding said.
  • fan renamed to bum
  • ss-SZ renamed to ss -- this was just a result of us having inconsistent data labels.
  • cjk merged into the gil dataset
  • bjj merged into the awa dataset

Canaries

Canaries are provided in separate canaries folder. Canaries are organized into three directions: monolingual hosts canaries designed for the MADLAD-400 monody data, multiway for the multiway data, and generic the generic canaries generated only from the model's vocabulary.

  • Monolingual: Canaries here are organized by the language the canary was generated from. This corresponds exactly to the translate_copy setting in the paper, where the source and target language match.

  • Multiway: Canaries here are organized in one of two fashions. to_XX indicates canaries organized by the target language (and where the source language could be any language). XX-XX indicates the canaries (interleaved_both and interleaved_mislabeled_both) designed for a specific pair of languages.

Within each subdirectory above, canaries are into separate files named by the canary type. There is always only a single file for each canary type. The generic folder contains within it the four canary types.

Canaries can be mixed in with normal training data to then be analyzed post-hoc to training

References

Raffel, Colin, et al. "Exploring the limits of transfer learning with a unified text-to-text transformer." J. Mach. Learn. Res. 21.140 (2020): 1-67.

Contact

Please reach out to {snehakudugunta, icaswell}꩜google.com. For questions about the canaries, reach out to cchoquette@google.com

License

This data is released with the CC-BY-4.0 license.

Detailed notes from the audit

Here are the notes on all languages, along with the number of documents found, and the final decision made with respect to including the language in this dataset.

Lang. note N decision
en ok 1838712272 keep
ru ok 402458746 keep
es good 250906994 keep
de ok 225111495 keep
fr ok 218863911 keep
it ok 126406256 keep
pt ok 124207090 keep
pl ok 90908786 keep
nl ok 86594116 keep
tr ok 56417359 keep
vi ok 54988654 keep
cs ok 38254671 keep
id ok 37979244 keep
ro ok 35397563 keep
sv ok. Also the last 35153050 keep
: : language (suz) is "ok : : :
: : bible" : : :
hu ok 29677075 keep
uk ok 24968305 keep
fa idk ask a farsi speaker; 23138888 keep
: : ALI: OK : : :
ja ok a little en mixed in 21818123 keep
el ok 20932239 keep
fi ok 20433664 keep
da ok 17865888 keep
th ok 17439979 keep
no ok 14864710 keep
bg ok 12755329 keep
ko ok 12653878 keep
ar good 12411641 keep
sk ok 11857945 keep
ca ok 9477390 keep
lt ok 8748025 keep
iw ok 7194574 keep
sl ok 6310419 keep
et ok 5542933 keep
lv ok 5007982 keep
hi ok some porn 4512205 keep
sq good 3622957 keep
az good 3256331 keep
hr ok 2841400 keep
ta ok 2594191 keep
ms ok 2337672 keep
ml ok 2072605 keep
sr ok 2010607 keep
kk ok 1810963 keep
te ok a lot of weirdly low 1682441 keep
: : quality looking content : : :
: : like commerce : : :
mr ok fix virama 1673848 keep
is ok 1560913 keep
bs good 1362582 keep
mk ok 1358293 keep
gl ok 1253170 keep
eu ok 1155671 keep
bn ok 1138848 keep
be ok 1092785 keep
ka ok 936497 keep
fil ok more bible than 901507 keep
: : expected for such a : : :
: : major language : : :
mn ok mongolian cyrillic 879878 keep
af good 868671 keep
uz ok some cyrllic noise 669909 keep
gu ok 659727 keep
kn ok 657846 keep
kaa ok cyrllic 586361 keep
sw ok 537847 keep
ur ok 467236 keep
ne ok 453349 keep
cy ok; was terrible before 430719 keep
: : filtering short docs : : :
hy ok 397523 keep
ky ok 367577 keep
si good 349220 keep
tt good plus some 346927 keep
: : nonunicode misrendered : : :
: : PDF : : :
tg good 328194 keep
la ok some broken chars 319178 keep
so good 293218 keep
ga ok some en noise 285999 keep
km ook 285740 keep
mt ok 265388 keep
eo ok; likely a lot of Mt 259971 keep
ps ok 252888 keep
rw ok 226466 keep
ku ok 218850 keep
lo ok many entities in 215982 keep
: : latin script : : :
fy ok plausible but i bet 210025 keep
: : there is a lot of nl in : : :
: : there : : :
ha ok 173485 keep
my filter noise and en fix 172401 keep
: : virama : : :
dv good 167179 keep
pa ok 150588 keep
ckb ok 148870 keep
lb ok 145988 keep
mg ok some bible jw 115387 keep
ht ok 110443 keep
ug ok 106549 keep
am good 106301 keep
or ok 100530 keep
fo good 97754 keep
gd ok 94275 keep
ba ok 90318 keep
tk ok; a few weird docs 82495 keep
mi ok 79509 keep
hmn ok 75213 keep
grc ok some bible 70730 keep
jv ok 69473 keep
ceb ok 66164 keep
sd good 65858 keep
yi ok 64949 keep
kaa-Latn ok urls are .ru or .kz 61169 keep
sn ok 60196 keep
co ok;l i suspect lots of 55387 keep
: : MT : : :
su good 54968 keep
pap ok 54498 keep
ig ok 54410 keep
zu good 53809 keep
xh ok 53672 keep
sm ok 52614 keep
ny ok 52244 keep
yo ok 52067 keep
cv good 47318 keep
el-Latn good; a lot of old 46428 keep
: : content! : : :
kl ok 46027 keep
haw ok scam tv products 45670 keep
gsw wtf is happening here; 42712 keep
: : keep with disclaimer; : : :
: : STILL BOILERPLATE : : :
tet good ; actually a lot of 40367 keep
: : fun data! : : :
st ok 40360 keep
lus ok 36437 keep
oc ok 36379 keep
as good 33825 keep
rm ok 33805 keep
br ok after shortfilter 33219 keep
sah ok 29169 keep
hi-Latn filter porn this is half 26723 keep
: : porn : : :
se good 23872 keep
cnh good, some local news! 21556 keep
: : not sure if WL : : :
om ok 18895 keep
ce ok 14968 keep
udm ok 13376 keep
lg ok lot of 13030 keep
: : www.bukedde.co.ug in : : :
: : this : : :
os ok 12623 keep
nv ok 12578 keep
kha ok 12070 keep
ilo ok some bible 11754 keep
ctd-Latn ok; from some local 11629 keep
: : news? : : :
vec very noisy has wiki from 11108 keep
: : other langs and .it : : :
: : websites so not sure if : : :
: : vec : : :
hil ok some en boilerplate 10564 keep
tyv ok fun stuff plus some 9083 keep
: : russian noise i think : : :
iba ok jw data 7638 keep
ru-Latn ok 7523 keep
kbd ok many .ru 7486 keep
ti ok; poor tigray 7288 keep
sa ok 7117 keep
av good 6331 keep
bo needs some serious 6226 keep
: : script filtering. but : : :
: : there is some ok data in : : :
: : there. : : :
zza good 6019 keep
ber-Latn ok 5612 keep
otq ok 5554 keep
te-Latn great good text....but 5305 keep
: : mostly pornographic : : :
bua ok 5264 keep
ts good 5198 keep
cfm ok mostly from 4858 keep
: : chinland.co : : :
tn good 4821 keep
krc ok 4815 keep
ak good; much but not all 4768 keep
: : bible : : :
meo ok mostly blogs 4655 keep
chm ok; fyi watch out for 4653 keep
: : yandex translationese : : :
to good ; news bible 4612 keep
: : government : : :
ee good; mostly religious 4536 keep
nso ok 4422 keep
ady good 4206 keep
rom bible 4187 keep
bho mostly from anjoria.com. 4121 keep
: : Looks like valid : : :
: : Bhojpuri. : : :
ltg ok mostly www.lakuga.lv 4120 keep
fj ok 3976 keep
yua ok 3965 keep
gn ok some broken 3858 keep
: : characters some bible : : :
az-RU good; a lot of JW 3781 keep
ln ok bible jw 3325 keep
ada good; bible; likely 3095 keep
: : mixed with gaa : : :
myv maybe has .ru urls 3095 keep
bik ok. keep in mind the bik 3092 keep
: : vs bcl issue. : : :
tlh ok, but why tf are there 3054 keep
: : websites inklingon? all : : :
: : MT ? : : :
kbp not sure if right script 3036 keep
: : wiki says latin : : :
war ok but v sus. Pls filter 2928 keep
: : out wikipedia : : :
wa ok lots of wiki stuff 2772 keep
bew mostly blogs. idk if 2677 keep
: : standard Indonesian or : : :
: : not : : :
rcf ok 2630 keep
ta-Latn good text .... but 2580 keep
: : pornographic : : :
kac ok 2567 keep
iu filter script some is en 2537 keep
: : rest is iu script : : :
ay good; mix of bible and 2505 keep
: : other news sources : : :
kum ok 2495 keep
qu ok 2449 keep
bgp almost all ur-Latn. 2427 keep
: : consider removing or : : :
: : renaming : : :
hif ok some en noise and 2358 keep
: : religious : : :
kw ok short boilerplate 2324 keep
: : bible wiki; ok some porn : : :
nan-Latn-TW ok 2285 keep
srn ok bible + jw 2281 keep
tly-IR deeply sus 2239 keep
sg ok jw 2106 keep
gom ok 2102 keep
ml-Latn ok some short docs 2071 keep
kj ok 2062 keep
ksd ok bible 2000 keep
dz ok; hidden parallel 1899 keep
: : text; maybe actually bo; : : :
: : mainly buddhist : : :
kv ok a lil boilerplate 1878 keep
: : vibes : : :
msi ok 1870 keep
ve ok mostly bible jw 1866 keep
zap ok JW. 1803 keep
zxx-xx-dtynoise BEAUTIFUL NOISE rename 1765 keep
: : but keep as beautiful : : :
: : xample. (was called : : :
: : "dty") : : :
meu ok bible 1728 keep
iso ok jw 1721 keep
ium filter out zh 1721 keep
nhe ok 1714 keep
tyz ok bible bu again i 1707 keep
: : think some mixeed : : :
: : dialects : : :
hui ok some bible 1680 keep
new ok 1634 keep
mdf ok some short docs 1609 keep
pag bible 1588 keep
gv filter short repetitive 1586 keep
: : sentences; still same : : :
: : but keep : : :
gag has 1-2 cyrillic 1572 keep
: : examples with small amts : : :
: : of arabic script noise : : :
ngu ok 1534 keep
quc bible 1526 keep
mam ok bible jw 1513 keep
min ok mostly wiki and bible 1474 keep
ho ok 1466 keep
pon bible 1462 keep
mrj ok 1447 keep
lu ok jw 1444 keep
gom-Latn ok very noisy ; some ok 1432 keep
: : stuff ; release with : : :
: : disclaimer : : :
alt ok 1422 keep
nzi ok 1371 keep
tzo ok bible + jw 1357 keep
bci ok bible 1329 keep
dtp ok; mostly from 1309 keep
: : www.newsabahtimes.com.my : : :
abt fine; bible 1305 keep
bbc ok 1274 keep
pck ok 1255 keep
mai ok mild amounts of en 1240 keep
: : noise : : :
mps ok bible 1239 keep
emp ok bible 1238 keep
mgh ok bible jw 1222 keep
tab idk plausibly ok 1202 keep
crh ok 1184 keep
tbz good mostly bible but 1126 keep
: : not all : : :
ss good mix of data ; 1089 keep
: : renamed from "ss" : : :
chk ok bible 1082 keep
bru ok; bible 1072 keep
nnb ok 1071 keep
fon ok mostly jw but not all 1065 keep
ppk bible 1063 keep
tiv ok jw 1063 keep
btx ok probably 1009 keep
bg-Latn ok 991 keep
mbt ok bible 969 keep
ace good; bible 966 keep
tvl ok jw 933 keep
dov ok bible + jw 923 keep
ach good; bible 915 keep
xal ok has .ru sites though 913 keep
cuk ok bible 899 keep
kos ok lds bible 881 keep
crs ok 873 keep
wo ok; mostly bible. 871 keep
bts ok; mostly bible 869 keep
ubu ok bible 846 keep
gym ok biblle 820 keep
ibb ok bible and repeated @ 818 keep
ape good; bible 814 keep
stq ok i think ? 809 keep
ang much noise but some good 803 keep
: : Old English in there! : : :
enq ok bible 793 keep
tsg much noise but somegood 789 keep
: : data too! : : :
shn mostly English 788 keep
: : boilerplate. filter by : : :
: : latin text before : : :
: : releasing : : :
kri ok boilerplate noise 786 keep
: : bible jw : : :
kek ok jw bible 782 keep
rmc ok 738 keep
acf good; bible 730 keep
syr good; practictitioners 716 keep
: : should keep dialect in : : :
: : mind. : : :
qub bible 705 keep
bm good 702 keep
tzh ok jw 702 keep
jiv ok bible 696 keep
kn-Latn filter en noise of 688 keep
: : karnatake govt websites : : :
kjh ok .ru domain 672 keep
yap ok 638 keep
ban ok bible 637 keep
tuc ok bible 635 keep
tcy good; mostly wikipedia; 632 keep
: : likely some konkani : : :
: : mixed in : : :
cab ok jw 629 keep
cak ok bible 617 keep
din ok after SD filter 611 keep
arn good; bible 593 keep
lrc ok 587 keep
gil empty; but merged in 586 keep
: : data in "cjk" : : :
gil this is all in gil 586 keep
: : (Kiribati). merged into : : :
: : "gil" : : :
rwo bible 572 keep
hus ok bible 569 keep
bum ok bible; but wrong 559 keep
: : language. Data is in : : :
: : Bulu, not Fang : : :
mak ok bible 555 keep
frp fair amount from 550 keep
: : wikipedia. : : :
seh ok jw 545 keep
twu ok bible, but also i 539 keep
: : think it's lots of mixed : : :
: : similar dialects : : :
kmb ok bible jw 538 keep
ksw ok bible 536 keep
sja ok bibe 527 keep
amu good; bible; crazy 511 keep
: : diacritics : : :
mad remove mostly short text 509 keep
quh bible 501 keep
dyu ok bible 483 keep
toj ok jw 452 keep
ch ok; not sure about WL 449 keep
sus hella sus jk ok bible 437 keep
nog ok 419 keep
jam ok bible 416 keep
gui ok bible 409 keep
nia ok 408 keep
mas ok some amount of bible 405 keep
bzj ok bible 404 keep
mkn ok bible 402 keep
lhu ok bible 377 keep
ctu ok bible 366 keep
kg ok bible jw 365 keep
inb ok bible 343 keep
guh ok bible 331 keep
rn bible 323 keep
bus ok; bible; about 50bzc 322 keep
mfe ok mostly bible maybe 320 keep
: : some french creole short : : :
: : doc noise : : :
sda ok bible 317 keep
bi good! fun! 311 keep
cr-Latn noise and lorem ipsom. 303 keep
: : But some ok Cree text. : : :
gor ok bible 303 keep
jac ok bible 303 keep
chr ok bible 301 keep
mh ok jw lds 296 keep
mni ok 290 keep
wal ok bible + jw 286 keep
teo ok bible 274 keep
gub ok bible 271 keep
qvi bible 266 keep
tdx ok jw 262 keep
rki ok 251 keep
djk ok; bible+jw 246 keep
nr ok 246 keep
zne ok jw 239 keep
izz ok bible 237 keep
noa ok 234 keep
bqc ok; bible 228 keep
srm ok; bible + jw 227 keep
niq ok 226 keep
bas ok; has some fun blog 216 keep
: : stuff! : : :
dwr ok; bible; mixed script 215 keep
guc ok bible 214 keep
jvn ok bible 213 keep
hvn ok religioous text 200 keep
sxn ok bible ; also wild 197 keep
: : diacritics : : :
koi ok 196 keep
alz good; bible 195 keep
nyu ok 195 keep
bn-Latn ok 191 keep
suz 186 keep
pau ok 185 keep
nij ok 183 keep
sat-Latn good! al from local news 183 keep
: : sources : : :
gu-Latn filter short en 179 keep
: : boilerplate and : : :
: : repetitive sentences : : :
msm ok bible 177 keep
maz ok bible jw 170 keep
qxr bible 153 keep
shp ok bible 150 keep
hne ok 146 keep
ktu ok bible jw 144 keep
laj ok bible 144 keep
pis bible 139 keep
mag ok fix virama issue 138 keep
gbm ok 137 keep
tzj ok bible 136 keep
oj ok 135 keep
ndc-ZW ok 132 keep
tks ok bible bu again i 127 keep
: : think some mixeed : : :
: : dialects : : :
gvl filter short boilerplate 126 keep
: : mostly bible : : :
knj ok bible 126 keep
awa all bible in awadhi 126 keep
: : (awa). Renamed from bjj : : :
spp ok bible 123 keep
mqy bible remove short docs 119 keep
tca ok bible + jw 117 keep
cce ok jw 116 keep
skr ok; some pnb mixed in 107 keep
kmz-Latn ok soome ar script noise 106 keep
dje ok; mostly but not all 100 keep
: : bible : : :
gof ok some bible 97 keep
agr good; bible 93 keep
qvz bible 88 keep
adh good; bible 87 keep
quf bible 86 keep
kjg ok bible 84 keep
tsc ok 82 keep
ber ok great! 79 keep
ify ok bible 79 keep
cbk ok bible 78 keep
quy bible 78 keep
ahk good; bible; crazy 77 keep
: : diacritics : : :
cac ok bible 77 keep
akb good; bible 71 keep
nut ok 67 keep
ffm ok bible; mixed fulfulde 65 keep
: : dialects; consider : : :
: : merging with ff : : :
taj ok bible 65 keep
ms-Arab ok mostly utusanmelayu 63 keep
: : website : : :
brx quite good! 62 keep
ann good; all from wikimedia 56 keep
: : incubator : : :
qup bible 53 keep
ms-Arab-BN ok not sure if same as 46 keep
: : ms-Arab : : :
miq ok 45 keep
msb ok bible 41 keep
bim good; bible 40 keep
raj ok 40 keep
kwi ok bible 37 keep
tll ok jw 37 keep
trp good ; lots of random 36 keep
: : stuff : : :
smt ok bible but lots of 34 keep
: : different bibles! : : :
mrw ok 29 keep
dln ok bible 28 keep
qvc bible 27 keep
doi ok actually nice! 26 keep
ff ok after shortfilter 26 keep
zh very noisy 19850947 keep (filtered)
zh-Latn poor quality 602 remove
rhg-Latn remove 10302 remove
ja-Latn remove maybe low quality 7516 remove
: : short and repeated : : :
pam remove 2773 remove
za revisit after 1700 remove
: : shortfilter : : :
ar-Latn terrible, 0% orrect, 1520 remove
: : remove : : :
mnw remove en noise and 1100 remove
: : boilerplate : : :
fip ok jw ; but wrong 729 remove
: : language. mostly : : :
: : Mambwe-Lungu and Bemba, : : :
: : as well as Fipu (mgr+bem : : :
: : vs. fip) : : :
el-CY bad; not Cypriote 537 remove
luz terrible; remove 354 remove
cni ok; bible; lots of mixed 261 remove
: : in content in : : :
: : not,cob,cpc,arl : : :
apd-SD terribly questionable; 227 remove
: : probably remove : : :
mey mostly short and noisy 127 remove
: : borderline : : :
awa OK; should be used with 126 remove
: : caution and suspicion : : :
mtq remove short doc 111 remove
: : repetitive : : :
mel remove noisy en 103 remove
mr-Latn remove mostly porn and 91 remove
: : short docs : : :
srr remove ; english 91 remove
: : boilerplate : : :
en-Cyrl ok ... some fr-Cyrl too 90 remove
: : and maybe others : : :
en-Arab remove 79 remove
syl idk maybe ok ? 61 remove
jax filter mostly 58 remove
: : text.medjugorje.ws : : :
: : boilerplate : : :
xmm very noisy lots of dj 58 remove
: : tiktok and peppa pig : : :
: : repeated : : :
shu quite questionable. prob 53 remove
: : remove : : :
ks ok shorter docs 51 remove
gyn remove boilerplate and 45 remove
: : porn : : :
aa some pretty bad data but 32 remove
: : also some good data. : : :
: : filter on "Woo" (case : : :
: : sensitive) : : :
sjp terible; probably 31 remove
: : remove; check again : : :
: : after short filter : : :
abs all short nonsense 24 remove
: : remove : : :
mui remove short docs 23 remove
mdh filter porn short text 22 remove
: : and repetitive : : :
: : boilerplate : : :
noe ok 22 remove
sxu rvisit after shortfilter 22 remove
bhb-Gujr bad. remove. all junk 20 remove
: : gu. : : :
yaq remove 20 remove
prk ok 18 remove
cgg rather noisy but 17 remove
: : potentialy ok. not sure : : :
: : if WL or not : : :
bto bad; remove unless short 16 remove
: : filter keeps enough : : :
ayl terrible 13 remove
pa-Arab ok 13 remove
bmm terrible. filter on 11 remove
: : short and reevaluate : : :
mfb remove short boilerplate 11 remove
mtr ok fix virama remove en 11 remove
: : noise : : :
pmy remove 11 remove
skg terrible; remove 11 remove
ymm remove 11 remove
xnr ok maybe fix virama 9 remove
: : though it seems fine : : :
kjb ok bible 8 remove
azg short noise; bible 7 remove
bgz idk maybe ok but 7 remove
: : probably bad : : :
ctg probably terrible 7 remove
: : probably remove : : :
nyo ok 7 remove
mdy ok bible 6 remove
syl-Latn revist or remove after 6 remove
: : shortfilter : : :
xog ok bible and stories 6 remove
cyo terrifying noise; remove 4 remove
kfy filter virama issue 4 remove
nd ok 4 remove
rwr remove 4 remove
tuf ok bible 4 remove
clu ok bible 3 remove
ng ok 3 remove
zyj deeply bad data .. 3 remove
: : revisit after : : :
: : shortfilter : : :
rkt ok 2 remove
bgc super sketch. Remove 1 remove
: : unless short doc filter : : :
: : leaves some. remove : : :
dcc remove 1 remove
ff-Adlm good 1 remove
gju remove short boilerplate 1 remove
max remove short some ru 1 remove
mwr filter short docs fix 1 remove
: : virama : : :
trw sus; remove 1 remove
vkt 1 doc remove 1 remove
gjk empty remove 0 remove
bfy very bad. remove unless 0 remove
: : it looks better after : : :
: : filtering short docs; : : :
: : remove : : :
nyn ok 0 remove
sgj remove 0 remove

A few comments too long to fit in the table above:

  • alt: WAIT THIS IS AMAZING IT IS ACTUALLY ALTAI! e.g. from urls like https://altaicholmon.ru/2020/02/28/jarashty-la-jajaltany-jarkyndu-lekeri/
  • tly-IR: They all look like boilerplate content, e.g., list of keywords/search queries used to bump page ranking in search results. Not any useful material for translation. Remove.
  • zap: pls note that at least some Zapotec speakers tend to view it as one language, not as a million dialects like ISO does. However, some are certainly mutually unintelligible, complicating the matter.
  • zh-Latn: The biggest problem is that several examples are not in Latin Chinese (i.e., romanization in my understanding) but in English or mixed English and Chinese. For those data in Latin Chinese, their quality seems to be good.
  • zh: Many examples are porn-related, particularly those very long documents. Also, there are some examples of traditional Chinese.

Final Dataset information

The number of documents, sentences, tokens, characters, and bytes for the noisy and clean splits of the data. Note that the "toks" field below uses whitespace for tokenization, so is not appropriate for non-whitespace-separating languages like Chinese (see section above). Note that the english subset in this version is missing 18% of documents that were included in the published analysis of the dataset. These documents will be incoporated in an update coming soon.

BCP-47 docs (noisy) docs (clean) sents (noisy) sents (clean) toks (noisy) toks (clean) chars (noisy) chars (clean) clean noisy
total* 7.2B 3.7B 133.1B 97.5B 4.6T 2.6T 30.6T 16.0T 11.4 T 6.3 T
en* 3.0B 1.5B 71.1B 45.4B 2.0T 1.3T 12.3T 7.6T 2.6 T 4.3 T
ru 823M 402.5M 823M 12.4B 416.5B 240.9B 3.1T 1.8T 832.9 G 1.4 T
es 476.4M 250.9M 8.3B 4.5B 325.7B 170.4B 2.1T 1.1T 380.9 G 747.5 G
de 478.6M 225.1M 11.5B 6B 299.5B 139.6B 2.2T 1T 370.6 G 815.5 G
fr 384.2M 218.9M 7.9B 5B 307.1B 165.2B 2T 1T 370.4 G 699.1 G
it 238.9M 126.4M 4.5B 2.5B 180.1B 83.6B 1.2T 553.1B 198.4 G 429.6 G
pt 209.2M 124.2M 4B 2.4B 123.2B 79.2B 791.5B 499.8B 183.1 G 289.6 G
pl 145.1M 90.9M 3.3B 2.4B 68.9B 49.2B 505B 356.4B 140.7 G 202.5 G
nl 134.5M 86.6M 134.5M 2.3B 104.4B 51.6B 698.5B 334.5B 118.2 G 247.5 G
tr 107M 56.4M 107M 1.2B 41.9B 25B 328.8B 198.9B 73.7 G 123.9 G
vi 92.8M 55M 1.6B 1B 71.5B 48.7B 342B 228.8B 88.8 G 133.9 G
cs 72.1M 38.3M 1.7B 1B 40.8B 22.1B 272.2B 147.9B 62.1 G 112.7 G
id 120.9M 38M 2.2B 747.5M 60.4B 20.2B 443B 148.3B 48.5 G 148.7 G
ro 60.8M 35.4M 60.8M 746.4M 37.1B 22.9B 244.1B 148.2B 55.5 G 90.3 G
sv 65.2M 35.2M 65.2M 1B 62.1B 23.9B 422.6B 153.7B 57.0 G 149.9 G
hu 47.6M 29.7M 1.3B 806.3M 29.8B 17.8B 223.6B 134.9B 53.5 G 86.8 G
uk 46.6M 25M 1B 599.9M 21.6B 12.8B 164.2B 95.2B 45.1 G 75.8 G
fa 58.1M 23.1M 920.6M 493.5M 40.6B 18.4B 220.4B 96.7B 43.4 G 97.4 G
ja 23.3M 21.8M 326M 321.6M 10.9B 10.9B 133.3B 132.2B 98.7 G 99.7 G
el 52.4M 20.9M 808M 445.4M 25B 12B 173.2B 80.9B 37.9 G 80.8 G
fi 35.8M 20.4M 1B 650.3M 23.8B 11.5B 202.2B 101.1B 37.6 G 74.1 G
zh 29.3M 19.9M 492.3M 298.8M 19.2B 10B 333B 142.3B 109.9 G 191.8 G
da 38.5M 17.9M 1.1B 508M 37.7B 13B 252B 83.1B 29.4 G 89.5 G
th 19M 17.4M 19M 385.8M 8.9B 8.9B 118.6B 117.6B 57.6 G 58.2 G
no 34.7M 14.9M 34.7M 498.7M 46.6B 11.8B 305.6B 74.8B 27.3 G 109.8 G
bg 27.2M 12.8M 599.4M 360.3M 14.4B 8.8B 95.6B 57.8B 26.0 G 42.8 G
ko 19.7M 12.7M 628.6M 471.8M 13.3B 9.3B 65.9B 43.8B 34.2 G 49.1 G
ar 67.6M 12.4M 876.6M 182.6M 39B 7.1B 243B 43.2B 20.9 G 115.9 G
sk 23.2M 11.9M 487.9M 300.6M 11.3B 6.7B 77.8B 45.7B 18.8 G 31.9 G
ca 17.9M 9.5M 258.6M 153M 8.9B 5.6B 56.5B 34.6B 12.6 G 20.8 G
lt 15.3M 8.7M 374M 256.9M 7.5B 5.3B 58.6B 41.3B 15.7 G 22.3 G
he 14.1M 7.2M 302.2M 196.8M 9.2B 5.2B 54.9B 30.5B 14.8 G 26.3 G
sl 12M 6.3M 316M 180M 6.9B 4.5B 47.8B 30.5B 11.5 G 18.0 G
et 8.8M 5.5M 223.8M 176.3M 5B 3.6B 40.1B 28.7B 10.7 G 15.0 G
lv 8.4M 5M 186.1M 138.5M 4.8B 3.2B 36.7B 23.9B 9.1 G 13.8 G
hi 9.9M 4.5M 254.4M 152M 7.4B 3.8B 39.9B 20.1B 9.9 G 19.7 G
sq 5.5M 3.6M 5.5M 56.1M 2.7B 2.1B 17B 12.7B 4.8 G 6.6 G
az 5.2M 3.3M 90.3M 70.9M 2.1B 1.5B 16.3B 11.9B 4.5 G 6.3 G
hr 23M 2.8M 476.6M 53M 12.6B 1.4B 85.1B 9.6B 3.7 G 33.5 G
ta 5.6M 2.6M 122.5M 81.9M 2.1B 1.1B 19.2B 10.6B 4.9 G 8.8 G
ms 14.1M 2.3M 14.1M 55.2M 8B 1.7B 58.8B 12.5B 4.0 G 20.4 G
ml 3.7M 2.1M 75M 52M 1B 603.3M 10.5B 6.3B 3.0 G 5.1 G
sr 4.7M 2M 4.7M 64M 2.7B 1.6B 18.6B 11B 5.1 G 8.7 G
kk 3.1M 1.8M 87.4M 59.1M 1.6B 1B 13.4B 8.6B 3.8 G 5.8 G
te 2.5M 1.7M 59M 46.4M 900.2M 618.5M 7.4B 5.1B 2.6 G 3.8 G
mr 2.9M 1.7M 2.9M 50M 1.2B 776.9M 8.7B 5.5B 2.8 G 4.4 G
is 2.9M 1.6M 73.7M 39.3M 2.1B 979.2M 14.9B 6.4B 2.5 G 5.9 G
bs 12.9M 1.4M 163.6M 9M 5.9B 490.9M 39.5B 3.3B 1.3 G 15.6 G
mk 2.9M 1.4M 41.3M 22.6M 1.3B 685.9M 9.1B 4.5B 2.0 G 4.0 G
gl 4.2M 1.3M 45.3M 18.8M 2.3B 748.4M 15.6B 4.8B 1.7 G 5.5 G
eu 2.1M 1.2M 41.7M 24.8M 827.5M 525.3M 6.9B 4.3B 1.5 G 2.4 G
bn 4.3M 1.1M 151.2M 38.6M 2.5B 645.7M 16.8B 4.3B 2.2 G 8.7 G
be 2M 1.1M 48.8M 31.3M 981M 632.9M 7.2B 4.6B 2.2 G 3.5 G
ka 3.1M 936.5K 53.7M 26.6M 1.2B 460.8M 10.3B 3.8B 1.9 G 5.0 G
fil 4.2M 901.5K 67.4M 19.2M 2.2B 741.7M 14.6B 4.7B 1.5 G 5.0 G
mn 2.2M 879.9K 43.3M 24M 1.1B 487.5M 7.9B 3.5B 1.6 G 3.5 G
af 2.9M 868.7K 51.9M 30M 1.7B 795M 11.8B 4.8B 1.8 G 4.2 G
uz 1.4M 669.9K 25.7M 17.5M 605.9M 388.3M 5.2B 3.3B 1.1 G 1.9 G
gu 1.3M 659.7K 28.9M 18.1M 634.4M 345.9M 3.9B 2.1B 1.1 G 2.0 G
kn 1.6M 657.8K 32.9M 19.2M 546.4M 258.6M 4.6B 2.2B 1.1 G 2.3 G
kaa 1.1M 586.4K 19.8M 13.3M 455.9M 269M 3.8B 2.2B 990.2 M 1.6 G
sw 1.3M 537.8K 1.3M 9.5M 660.7M 345.8M 4.6B 2.4B 826.1 M 1.6 G
ur 967.2K 467.2K 29M 18.4M 1B 562.5M 5.2B 2.7B 1.2 G 2.4 G
ne 876.4K 453.3K 876.4K 20.4M 585M 345.3M 3.9B 2.2B 1.1 G 1.9 G
cy 4.9M 430.7K 68.3M 7.4M 3.6B 275.6M 26.4B 1.7B 609.5 M 10.0 G
hy 2M 397.5K 31.1M 9.9M 1B 190.9M 8.1B 1.5B 678.9 M 3.6 G
ky 751.1K 367.6K 14.3M 9.6M 303.4M 181.6M 2.5B 1.4B 665.1 M 1.1 G
si 788K 349.2K 22.1M 16M 507.3M 293.3M 3.4B 1.9B 1023.6 M 1.8 G
tt 2.1M 346.9K 60.2M 8.6M 1B 135M 12.1B 1B 494.1 M 4.6 G
tg 789.2K 328.2K 789.2K 7.4M 363.8M 208.8M 2.6B 1.4B 635.7 M 1.1 G
la 2.9M 319.2K 85.7M 13.8M 1.1B 218.4M 8.2B 1.5B 550.6 M 2.9 G
so 729.2K 293.2K 729.2K 3.1M 294.8M 146.3M 2.1B 992.4M 350.8 M 746.2 M
ga 5.3M 286K 31.7M 6.9M 4.2B 229.3M 30.6B 1.4B 500.7 M 9.8 G
km 297.8K 285.7K 5M 5M 53M 52.6M 1.1B 1.1B 566.2 M 570.0 M
mt 1.2M 265.4K 1.2M 5.6M 390.4M 171.5M 3.2B 1.3B 467.4 M 1.1 G
eo 1.4M 260K 33.9M 9.3M 745.1M 253.1M 5.5B 1.7B 627.6 M 1.9 G
ps 429.9K 252.9K 5.1M 3.6M 293.9M 177.5M 1.4B 848.9M 403.5 M 682.9 M
rw 681.8K 226.5K 681.8K 1.9M 225M 99.8M 1.7B 749.1M 264.8 M 702.4 M
ku 671.9K 218.9K 10.7M 4.9M 305.3M 143.8M 2.1B 849.9M 335.3 M 791.9 M
lo 229.1K 216K 2.9M 2.8M 41.7M 41.1M 706.9M 697.6M 365.3 M 370.8 M
fy 1.7M 210K 12.1M 3.7M 506.9M 94M 3.7B 592.3M 223.0 M 1.2 G
ha 443.9K 173.5K 4.5M 2.4M 206.5M 109.3M 1.3B 630.2M 219.0 M 478.1 M
my 176.5K 172.4K 176.5K 10.1M 96.6M 96.3M 1.3B 1.3B 648.8 M 650.4 M
dv 264.4K 167.2K 4.3M 3.5M 92.8M 64M 877.3M 603.1M 238.3 M 343.2 M
pa 368.2K 150.6K 368.2K 6M 306M 152.8M 1.6B 797.1M 414.1 M 857.6 M
ckb 622.7K 148.9K 5.6M 2.5M 312.7M 83.3M 2.2B 572.7M 265.0 M 1011.1 M
lb 7.6M 146K 47.1M 3.4M 7.5B 85M 58.4B 575.5M 218.4 M 22.2 G
mg 295.2K 115.4K 4.5M 2.6M 189.4M 75.5M 1.3B 548.5M 179.0 M 429.3 M
ht 425.6K 110.4K 6.7M 2.6M 163M 84.3M 994.5M 461.5M 168.2 M 361.5 M
ug 227.1K 106.5K 4.5M 3.1M 122.9M 62.7M 998.5M 504.6M 233.1 M 449.9 M
am 245.2K 106.3K 7.1M 5.3M 157M 95.2M 869.9M 509M 345.5 M 539.4 M
or 139.6K 100.5K 139.6K 3.1M 66M 47.3M 437.2M 309.5M 160.3 M 228.1 M
fo 382.9K 97.8K 3.9M 1.8M 136.5M 48.9M 923.3M 314.9M 122.0 M 328.8 M
gd 206K 94.3K 3.7M 2.4M 127.6M 84.5M 812M 526M 173.4 M 276.6 M
ba 372.4K 90.3K 9.3M 2.6M 101M 42.1M 766.5M 320.7M 154.8 M 352.4 M
tk 180.2K 82.5K 180.2K 1.8M 65.4M 43.3M 575.2M 369M 131.3 M 221.6 M
mi 711.9K 79.5K 5.9M 1.9M 262.5M 73.5M 1.6B 371.9M 120.2 M 539.1 M
hmn 241.3K 75.2K 3.5M 1.9M 192.1M 80.2M 1.2B 408.8M 124.3 M 366.0 M
grc 364.8K 70.7K 13.7M 2.8M 298.6M 65.3M 2B 417.8M 217.7 M 1.0 G
jv 999.5K 69.5K 13M 2M 302.3M 52.1M 2.3B 376.1M 130.9 M 797.8 M
ceb 617.5K 66.2K 6.7M 1.6M 225M 58.2M 1.5B 357.7M 116.2 M 451.4 M
sd 115.6K 65.9K 115.6K 2.4M 112.6M 77.8M 561M 380.4M 182.3 M 267.1 M
yi 160.6K 64.9K 3.3M 1.9M 129.1M 53.9M 838.4M 352.6M 146.0 M 350.8 M
kaa_Latn 375.2K 61.2K 3.6M 1.3M 375.2K 61.2K 1.5M 209.5K 86.2 M 264.6 M
sn 3.1M 60.2K 3.1M 1.2M 1.3B 31.6M 10.6B 266M 92.5 M 3.2 G
co 546.7K 55.4K 6.1M 1.3M 172.6M 43.6M 1.1B 265.5M 98.8 M 386.8 M
su 336.6K 55K 336.6K 1.6M 154M 39.5M 967.2M 286.7M 100.7 M 308.5 M
pap 259.1K 54.5K 259.1K 1.4M 183.9M 41.1M 1.4B 229.9M 83.5 M 451.4 M
ig 130.4K 54.4K 2.1M 1.4M 129.2M 45.7M 846.1M 251.4M 93.0 M 178.9 M
zu 372.3K 53.8K 3.8M 1.2M 148.4M 27.2M 1.2B 257.4M 89.6 M 374.7 M
xh 310.9K 53.7K 2.9M 1.4M 81.6M 31.2M 749.5M 287.3M 100.0 M 319.1 M
sm 137.8K 52.6K 1.9M 1.3M 100.9M 53.7M 607.9M 276.3M 88.6 M 184.5 M
ny 181.6K 52.2K 181.6K 1.5M 80.6M 34.8M 611.2M 277.5M 91.8 M 209.8 M
yo 115K 52.1K 2M 1.2M 76.6M 46.3M 415.6M 239M 89.2 M 157.8 M
cv 599.4K 47.3K 12M 1.6M 169.6M 22.2M 1B 168.9M 82.1 M 413.6 M
el_Latn 497.3K 46.4K 11.3M 1.7M 497.3K 46.4K 2.3M 162.8K 196.8 M 571.1 M
kl 85.9K 46K 2.1M 1.5M 32.3M 22.3M 403.9M 279.1M 84.2 M 126.1 M
haw 310.4K 45.7K 7.1M 1M 141M 43.3M 892M 214.2M 69.9 M 271.2 M
gsw 7.6M 42.7K 64.5M 1M 5B 22.3M 42.3B 149.2M 53.8 M 13.5 G
tet 291K 40.4K 1.9M 475.7K 240.6M 22.8M 1.6B 152.3M 51.2 M 455.4 M
st 96.8K 40.4K 96.8K 1.1M 65M 39.8M 381.5M 226.9M 74.0 M 127.0 M
lus 91.5K 36.4K 1.4M 863.5K 53M 31.3M 298.3M 167.3M 60.1 M 107.0 M
oc 2.4M 36.4K 2.4M 1.6M 887.6M 26.7M 6.7B 177.6M 58.7 M 1.9 G
as 53.9K 33.8K 2.4M 1.7M 41.4M 27.9M 275.8M 182.1M 95.8 M 146.1 M
rm 238.1K 33.8K 238.1K 603.4K 59.2M 15.8M 391M 100.2M 34.6 M 133.1 M
br 705.4K 33.2K 7.8M 731.7K 646.8M 21M 3.7B 125.4M 46.2 M 1.2 G
sah 1.3M 29.2K 1.3M 1.2M 283.7M 17.6M 2.2B 148.2M 68.3 M 852.3 M
hi_Latn 1.2M 26.7K 22.6M 1.2M 1.2M 26.7K 5.3M 98.9K 53.5 M 1.7 G
se 54.3K 23.9K 879.5K 493.3K 17.7M 10M 148.4M 84.6M 31.1 M 56.6 M
cnh 44.4K 21.6K 688.6K 406.9K 21.6M 12.5M 110.8M 63M 22.1 M 39.6 M
om 846.1K 18.9K 846.1K 469.8K 238M 11.2M 1.9B 88.5M 30.4 M 881.5 M
ce 59.3K 15K 991.1K 460.1K 17.8M 9.6M 130.6M 67.8M 31.1 M 60.2 M
udm 67.1K 13.4K 942.7K 510.3K 14M 7.4M 106M 55.5M 26.3 M 49.2 M
lg 61.1K 13K 510.9K 166.1K 21.4M 6.1M 160.7M 48M 17.3 M 56.7 M
os 172.1K 12.6K 172.1K 359.3K 27.1M 6.9M 233.5M 50.1M 23.1 M 87.7 M
nv 17.1K 12.6K 17.1K 86.5K 3.1M 1.1M 24.8M 9.1M 2.0 M 7.9 M
kha 37.8K 12.1K 235.5K 75.2K 15.8M 6M 88.6M 30.2M 9.8 M 27.3 M
ilo 69.8K 11.8K 889.2K 365.1K 26.7M 9M 187.9M 59.4M 20.6 M 64.0 M
ctd_Latn 23.3K 11.6K 575.6K 382.2K 23.3K 11.6K 90.7K 41K 21.5 M 35.1 M
vec 1.1M 11.1K 10M 209.7K 284.7M 7.8M 1.8B 43.8M 17.7 M 625.0 M
hil 126.8K 10.6K 1.1M 379.7K 43.9M 9.2M 293.5M 57.2M 18.5 M 95.2 M
tyv 61.6K 9.1K 596.6K 268.3K 9.9M 4.7M 80.2M 38.5M 16.7 M 36.6 M
iba 34K 7.6K 326.9K 126.1K 37.8M 4.8M 251.4M 30.5M 10.0 M 61.3 M
ru_Latn 346.3K 7.5K 346.3K 239.1K 346.3K 7.5K 1.5M 27.7K 14.9 M 452.3 M
kbd 154.7K 7.5K 1.4M 257.2K 31.9M 4.4M 321.4M 36.8M 16.8 M 209.6 M
ti 20.8K 7.3K 20.8K 481.3K 18.2M 8.8M 95.4M 44.6M 30.9 M 63.6 M
sa 154.3K 7.1K 154.3K 1.1M 70M 9.9M 512.5M 88.8M 44.9 M 236.6 M
av 107.6K 6.3K 806.1K 190.1K 15.5M 3.4M 129M 30.2M 12.8 M 56.0 M
bo 6.2K 6.2K 1.1M 1.1M 3.4M 3.4M 88.7M 88.7M 40.7 M 40.7 M
zza 370.1K 6K 3.3M 229.2K 87.7M 3.9M 617.3M 26.3M 10.0 M 234.1 M
ber_Latn 480.5K 5.6K 10.5M 169.4K 480.5K 5.6K 2.1M 18.9K 11.0 M 945.3 M
otq 17.6K 5.6K 17.6K 114.8K 10.2M 3.8M 65M 23.4M 7.7 M 22.8 M
te_Latn 236.6K 5.3K 4.4M 269.1K 236.6K 5.3K 1M 19.3K 11.4 M 254.3 M
bua 9.8K 5.3K 252K 144.6K 4.7M 2.7M 38M 21.7M 10.0 M 17.9 M
ts 34.7K 5.2K 34.7K 248.6K 39.6M 6.5M 377.2M 38.8M 12.2 M 99.5 M
cfm 9.1K 4.9K 199.6K 128.6K 6.2M 4M 32.9M 21.5M 7.4 M 11.6 M
tn 138.2K 4.8K 138.2K 174.4K 46M 5.5M 302.3M 29.2M 9.4 M 99.0 M
krc 359.5K 4.8K 2.3M 153.9K 50.2M 2.6M 369.5M 20.7M 9.1 M 139.9 M
ak 19.5K 4.8K 341.7K 210.2K 12.3M 4.7M 74.5M 24.8M 9.1 M 24.7 M
meo 790.7K 4.7K 16.5M 39K 478M 1.2M 3B 7.5M 3.1 M 1.2 G
chm 81.5K 4.7K 929.1K 179.7K 17.2M 2.9M 132.2M 21.3M 9.8 M 53.5 M
to 14.3K 4.6K 14.3K 149K 10.3M 5.7M 58.2M 29.9M 9.6 M 19.0 M
ee 14.1K 4.5K 353.6K 246.7K 9.7M 6.2M 67.9M 32.8M 11.8 M 23.3 M
nso 376.2K 4.4K 376.2K 188.4K 419.2M 5.3M 2B 28.2M 9.1 M 502.7 M
ady 74.9K 4.2K 446.8K 96.9K 8M 1.6M 67.9M 14.8M 6.4 M 30.6 M
rom 22.9K 4.2K 22.9K 76.1K 8.9M 2.6M 59M 15.9M 5.8 M 21.0 M
bho 13.6K 4.1K 306.2K 118.5K 7.1M 2.7M 37.6M 13.4M 7.4 M 20.6 M
ltg 13.1K 4.1K 213.7K 87.3K 4M 1.9M 29.2M 13.9M 5.6 M 11.7 M
fj 17K 4K 410K 164.1K 11.6M 5.2M 67.7M 28M 8.6 M 22.5 M
yua 10.4K 4K 141.6K 77.6K 5.2M 2.5M 36.8M 17.2M 5.7 M 12.4 M
gn 87.1K 3.9K 770.9K 162.6K 19.2M 2.7M 140.7M 20.8M 7.8 M 52.1 M
az_RU 6.5K 3.8K 231.8K 177.3K 6.5K 3.8K 24K 12.9K 10.3 M 15.1 M
ln 94.7K 3.3K 718.7K 139K 42.4M 3.4M 291.8M 21.5M 6.8 M 85.3 M
ada 6.5K 3.1K 291.5K 199.2K 7.5M 4.9M 38.9M 24.2M 8.6 M 13.9 M
myv 164.8K 3.1K 164.8K 130K 16M 1.7M 120.3M 13.8M 6.2 M 49.5 M
bik 44.8K 3.1K 376.7K 77K 14.8M 2.5M 102.3M 15.7M 5.3 M 34.0 M
tlh 516.9K 3.1K 516.9K 46.9K 221.3M 1.1M 1.4B 7.8M 2.7 M 554.2 M
kbp 5.9K 3K 247.9K 128.3K 5.6M 2.6M 30.8M 14.6M 5.7 M 12.4 M
war 1M 2.9K 114M 96.2K 612.1M 2.4M 3.5B 16.1M 3.7 M 1.2 G
wa 70.6K 2.8K 1.5M 127.2K 35.2M 3.6M 198.8M 20.4M 7.2 M 67.8 M
bew 311.1K 2.7K 10.4M 58.4K 212.4M 1.3M 1.4B 8.5M 3.1 M 547.1 M
rcf 21.6K 2.6K 21.6K 50.5K 4.9M 1.2M 30.2M 5.7M 2.1 M 11.4 M
ta_Latn 260.7K 2.6K 3.4M 142.7K 260.7K 2.6K 1.2M 9.1K 5.0 M 215.4 M
kac 5.9K 2.6K 109.2K 77.4K 5M 2.8M 26.6M 13.6M 4.3 M 8.0 M
iu 5.4K 2.5K 92.6K 53.1K 1.9M 907.4K 17.5M 8.3M 4.8 M 9.9 M
ay 8.1K 2.5K 196.7K 83.8K 3.9M 1.4M 34.5M 13.1M 4.5 M 12.7 M
kum 4.2K 2.5K 132.2K 89.7K 2.3M 1.6M 18.2M 12.4M 5.3 M 8.0 M
qu 149.7K 2.4K 1M 87K 26.7M 1.3M 200.6M 12.2M 4.0 M 68.3 M
bgp 355.7K 2.4K 5.6M 43.3K 186.1M 1.8M 1.1B 9.8M 3.1 M 377.5 M
hif 702K 2.4K 7.9M 124.7K 1.2B 3.2M 9.1B 19.1M 5.9 M 3.5 G
kw 176.9K 2.3K 1M 51.6K 53.1M 1.3M 327.8M 7.7M 2.8 M 89.2 M
nan_Latn_TW 7.4K 2.3K 7.4K 72.7K 7.4K 2.3K 28.3K 7.7K 4.8 M 15.4 M
srn 16.7K 2.3K 16.7K 139.5K 8M 3.4M 49.1M 17M 5.1 M 15.6 M
tly_IR 406.3K 2.2K 406.3K 18.2K 406.3K 2.2K 1.6M 8.6K 580.4 K 283.0 M
sg 4.2K 2.1K 154K 117.9K 4.6M 3.3M 22.6M 15.5M 4.6 M 6.8 M
gom 4.6K 2.1K 178.3K 108K 2.7M 1.4M 19.8M 10M 5.0 M 10.5 M
ml_Latn 260.8K 2.1K 3.5M 77.3K 260.8K 2.1K 1.1M 7.2K 3.5 M 277.7 M
kj 112.2K 2.1K 881.8K 22.6K 46.9M 877.3K 339.6M 6M 2.1 M 104.9 M
ksd 14.9K 2K 533K 78.6K 11.5M 2.1M 62.4M 10M 2.9 M 20.0 M
dz 1.9K 1.9K 191.7K 191.7K 1.1M 1.1M 22.7M 22.7M 10.0 M 10.0 M
kv 59.1K 1.9K 584.3K 88.8K 9.5M 1.2M 91.4M 9M 4.4 M 41.0 M
msi 686.7K 1.9K 686.7K 22.6K 414.8M 440.4K 2.6B 2.7M 1.1 M 1.0 G
ve 3.8K 1.9K 97.8K 79.4K 3.2M 2.1M 19M 11.7M 3.8 M 6.2 M
zap 5.5K 1.8K 202.3K 93.5K 4.2M 1.8M 26.4M 11.4M 4.0 M 9.6 M
zxx_xx_dtynoise 118.8K 1.8K 3.8M 49.3K 118.8K 1.8K 501K 6.6K 3.9 M 367.0 M
meu 5.9K 1.7K 232.1K 72.6K 4.2M 1.4M 27.2M 8.6M 2.6 M 9.1 M
iso 3.7K 1.7K 155.8K 111.5K 4.4M 2.7M 23M 13.7M 4.9 M 8.1 M
ium 100.3K 1.7K 6.2M 54.9K 48.4M 1.7M 314M 7.4M 2.6 M 124.0 M
nhe 3K 1.7K 3K 57.7K 1.9M 1.2M 15.6M 9.8M 2.7 M 4.8 M
tyz 8K 1.7K 454.8K 104.6K 7.5M 1.9M 46.3M 11.3M 3.8 M 16.0 M
hui 2K 1.7K 80.1K 74.7K 1.8M 1.7M 11.8M 10.9M 3.0 M 3.3 M
new 6.6K 1.6K 6.6K 85K 3.2M 1.4M 21.2M 8.8M 4.4 M 10.6 M
mdf 71K 1.6K 394.7K 45.1K 8.3M 670.1K 65.8M 5.5M 2.5 M 26.7 M
pag 49.6K 1.6K 49.6K 88.8K 13.8M 1.9M 92.9M 12M 3.9 M 29.2 M
gv 501.9K 1.6K 18.8M 26.9K 137.7M 996.2K 933.1M 6.2M 2.0 M 318.6 M
gag 33.9K 1.6K 491K 37K 10.2M 661K 84.9M 5.2M 2.1 M 32.6 M
ngu 3.8K 1.5K 3.8K 87.1K 2.7M 1.5M 21.4M 11.8M 3.6 M 6.7 M
quc 4.4K 1.5K 89.2K 41.2K 2.8M 1.1M 16.6M 6.4M 2.2 M 5.9 M
mam 23K 1.5K 446.3K 52.9K 9.8M 1.2M 70.4M 7.2M 2.6 M 30.7 M
min 28.2K 1.5K 500.9K 75.6K 10.2M 1.4M 70.5M 9.9M 2.6 M 21.1 M
ho 2K 1.5K 57K 47.8K 1.8M 1.3M 12.3M 7.8M 1.9 M 3.1 M
pon 5.7K 1.5K 167.8K 48.7K 3M 1.1M 18.3M 6.7M 2.1 M 6.1 M
mrj 97.1K 1.4K 97.1K 60.3K 14.5M 1.1M 100.6M 7.6M 3.6 M 40.8 M
lu 10.6K 1.4K 316K 112.1K 7.8M 2.3M 54.2M 15.4M 4.8 M 18.0 M
gom_Latn 231.1K 1.4K 4.1M 77.9K 231.1K 1.4K 1M 5.1K 3.6 M 240.6 M
alt 2.6K 1.4K 110.1K 65.9K 1.8M 1.1M 14.3M 8.7M 3.8 M 6.4 M
nzi 2.5K 1.4K 2.5K 71.8K 2.5M 1.7M 14.4M 9.4M 3.1 M 4.8 M
tzo 2.8K 1.4K 100.4K 75.7K 2.5M 1.7M 15.9M 10.6M 3.2 M 4.9 M
bci 7.4K 1.3K 124.8K 87.1K 5M 1.9M 32.8M 9M 3.1 M 9.4 M
dtp 4.6K 1.3K 51.2K 7.9K 1.9M 419.4K 12.7M 3M 1013.9 K 4.5 M
abt 1.6K 1.3K 122.7K 110.3K 1.5M 1.3M 9.6M 8.2M 2.2 M 2.7 M
bbc 72.3K 1.3K 718.3K 73.2K 21.7M 1.7M 151.3M 10.6M 3.6 M 47.9 M
pck 8.9K 1.3K 8.9K 69.7K 6.8M 2.1M 39.8M 11.5M 4.2 M 14.2 M
mai 54.3K 1.2K 1M 60.2K 24.6M 1.2M 156M 6.8M 3.6 M 67.1 M
mps 2.7K 1.2K 132.8K 71.9K 2.8M 1.6M 16M 8.7M 2.3 M 4.8 M
emp 3.6K 1.2K 106.4K 75.4K 1.9M 999.1K 14.5M 7.4M 2.4 M 4.9 M
mgh 5.5K 1.2K 151.8K 61.2K 2.8M 1.1M 24.1M 8.2M 2.8 M 8.3 M
tab 7.8K 1.2K 226.4K 26.8K 4.3M 538.9K 33.7M 4.4M 1.9 M 15.7 M
crh 5.1K 1.2K 170.9K 61.8K 2.4M 943K 18.8M 7.5M 3.4 M 8.9 M
tbz 5.1K 1.1K 128.7K 37.5K 3.5M 893.4K 22M 4.8M 1.9 M 10.2 M
ss 8.1K 1.1K 8.1K 30.4K 2.7M 568.3K 23.7M 5.5M 1.8 M 7.4 M
chk 2.8K 1.1K 98.8K 44K 2M 1M 12M 5.8M 1.8 M 4.0 M
bru 3K 1.1K 89.7K 48.2K 2.4M 938.1K 12.9M 4.8M 1.5 M 4.5 M
nnb 4.9K 1.1K 4.9K 70.2K 3.2M 1.2M 27.7M 9.1M 3.3 M 10.0 M
fon 5.3K 1.1K 222.9K 67.3K 6.9M 1.8M 34M 8.3M 3.1 M 14.8 M
ppk 2.6K 1.1K 85.8K 34.9K 1.9M 801.8K 13.2M 5.5M 1.6 M 4.3 M
tiv 3.8K 1.1K 3.8K 80.7K 3.7M 2.1M 20.4M 10.2M 3.2 M 6.0 M
btx 3.1K 1K 81.7K 43.9K 2M 907.5K 13.1M 5.9M 2.0 M 4.6 M
bg_Latn 200.4K 991 2.8M 25.5K 200.4K 991 927.1K 3.7K 1.7 M 143.6 M
mbt 1.6K 969 86K 45.4K 2.4M 1.3M 14.6M 7.5M 2.2 M 5.1 M
ace 65.5K 966 632.5K 32.5K 19.9M 1.1M 146.1M 7.4M 2.2 M 42.3 M
tvl 2.3K 933 72.9K 53.6K 2.5M 1.7M 12.6M 8.1M 2.4 M 3.8 M
dov 3.5K 923 129.8K 56.7K 2.6M 967.5K 20.7M 8M 2.6 M 7.1 M
ach 2K 915 63K 40.1K 1.6M 890.9K 9M 4.7M 1.6 M 3.0 M
xal 71.8K 913 498.5K 30.8K 8.5M 449.8K 64.7M 3.2M 1.5 M 24.4 M
cuk 4.1K 899 76.5K 34.3K 2M 469.9K 24.7M 4.6M 1.5 M 6.1 M
kos 2.2K 881 44.6K 27.8K 1.1M 780.1K 6.5M 4.2M 1.4 M 2.2 M
crs 7.6K 873 282.4K 40.1K 7.3M 1.2M 40.1M 6.8M 2.2 M 13.2 M
wo 36.4K 871 303.4K 25.4K 30.7M 850.7K 213.4M 4.5M 1.7 M 59.9 M
bts 3.2K 869 109.1K 29.1K 3.1M 663.3K 20.8M 4.2M 1.4 M 6.2 M
ubu 2.2K 846 113.5K 47.5K 2.3M 996.4K 15.9M 6.7M 1.9 M 4.7 M
gym 1.5K 820 73.7K 49.6K 1.6M 1.1M 10.3M 6.9M 2.0 M 3.2 M
ibb 74.1K 818 516.5K 36.3K 26.4M 776.1K 190.9M 4.9M 1.5 M 56.0 M
ape 7K 814 147K 56.1K 12.4M 881.5K 71M 5.8M 1.6 M 18.8 M
stq 111.9K 809 111.9K 27.7K 34.4M 600.4K 243.1M 3.8M 1.5 M 82.5 M
ang 66.5K 803 1.8M 86.7K 28.5M 1.7M 193M 9.8M 3.4 M 67.1 M
enq 7.1K 793 241.9K 39.1K 11M 718.8K 68.5M 4.8M 1.3 M 18.8 M
tsg 353.8K 789 353.8K 17.9K 158M 588.9K 1.1B 3.8M 1.0 M 309.9 M
shn 889 788 46.4K 46.2K 383.8K 378.5K 5.7M 5.7M 2.6 M 2.6 M
kri 39.1K 786 271.2K 38.8K 12.6M 995.2K 86.4M 5M 1.6 M 20.9 M
kek 3.2K 782 70.4K 38.4K 1.8M 709K 13.6M 4.4M 1.4 M 4.7 M
rmc 2.4K 738 2.4K 25.8K 1.3M 545.4K 7.9M 3.2M 1.1 M 2.9 M
acf 4.9K 730 81.9K 24.6K 2.1M 602.2K 11.6M 3M 1.1 M 4.7 M
fip 3.7K 729 165.6K 49K 3.5M 916.8K 25.7M 6.6M 2.1 M 8.6 M
syr 3.5K 716 326.4K 197.1K 4.6M 1.9M 31.5M 14M 6.1 M 13.9 M
qub 972 705 61K 51.1K 589.2K 455.5K 5.9M 4.4M 1.4 M 1.8 M
bm 21.9K 702 172.3K 24.5K 7.1M 583.1K 48.4M 3M 1.1 M 14.4 M
tzh 1.7K 702 41.7K 33.9K 1.5M 929.6K 9.3M 5.6M 1.6 M 2.6 M
jiv 1.7K 696 80.9K 32K 1.1M 418.9K 9.6M 3.5M 1.1 M 3.3 M
kn_Latn 72.9K 688 765.9K 10.1K 72.9K 688 328.1K 2.5K 430.8 K 61.4 M
kjh 1.5K 672 42.8K 28.7K 566.1K 379.2K 4.5M 3.1M 1.3 M 2.0 M
yap 1.9K 638 37.6K 19.5K 1.3M 661.4K 6.9M 3.3M 1.0 M 2.2 M
ban 8K 637 150.9K 16.3K 5M 499.7K 35.4M 3.6M 1.1 M 12.0 M
tuc 3.5K 635 193.2K 50.3K 2.9M 703K 17.2M 4.1M 1.2 M 5.7 M
tcy 10.7K 632 338.7K 37.1K 5.5M 432.6K 41.6M 3.3M 1.7 M 20.9 M
cab 1.2K 629 50.4K 37.5K 1M 690.9K 7.5M 5.1M 1.6 M 2.4 M
cak 1.2K 617 70.4K 32.6K 1.3M 730.1K 7.6M 4.2M 1.3 M 2.4 M
din 128.4K 611 885.8K 23.6K 31.6M 541.7K 210M 2.9M 1.1 M 64.3 M
zh_Latn 739.4K 602 10.7M 45.1K 739.4K 602 3.4M 2.3K 2.0 M 969.9 M
arn 2.4K 593 64.5K 26.2K 1.5M 541.9K 10.2M 3.7M 1.2 M 3.7 M
lrc 42.4K 587 351.9K 9K 17.3M 248.9K 85.3M 1.4M 646.9 K 37.5 M
rwo 938 572 938 45.5K 734.8K 590.4K 5.1M 4.2M 1.1 M 1.4 M
hus 825 569 26.5K 23.7K 733.4K 542.1K 4.4M 3.1M 967.6 K 1.3 M
bum 4.7K 559 103.8K 36.5K 3M 805.5K 18.8M 4M 1.3 M 6.1 M
mak 1K 555 32.5K 20.4K 761K 457.4K 6.1M 3.7M 1.1 M 2.0 M
frp 148K 550 3.5M 8.2K 71.2M 230.2K 535.4M 1.4M 518.3 K 129.7 M
seh 5.6K 545 68.8K 37.2K 2M 650.6K 14.9M 4.9M 1.5 M 4.4 M
twu 2.5K 539 109.9K 24.4K 2.4M 571.2K 14.2M 3.2M 1.0 M 4.8 M
kmb 1.3K 538 60.4K 36.9K 1.4M 810.8K 8.4M 4.6M 1.4 M 2.6 M
ksw 560 536 16.1K 16K 219.9K 218.8K 2.9M 2.9M 1.4 M 1.4 M
sja 1.3K 527 67.7K 24.9K 982.5K 459.3K 7.7M 3.4M 1.1 M 2.6 M
amu 1.8K 511 72K 25.2K 1.5M 443.3K 9.6M 3.2M 1.0 M 3.4 M
mad 103.8K 509 500.6K 18.5K 16.2M 386.7K 111.8M 2.8M 960.3 K 34.2 M
quh 1K 501 42K 29.9K 624.4K 396.8K 5.8M 3.7M 1.2 M 1.8 M
dyu 1.2K 483 55.8K 19.7K 1.2M 421.8K 5.7M 2M 665.5 K 1.9 M
toj 736 452 736 26.1K 691.2K 540.2K 4.3M 3.3M 1.0 M 1.3 M
ch 12.9K 449 147.5K 16K 8.9M 393.9K 63.5M 2.5M 906.8 K 10.0 M
sus 664 437 664 15.2K 648K 402.8K 3.7M 2.1M 674.0 K 1.0 M
nog 970 419 970 11K 330.3K 200.4K 2.6M 1.6M 714.0 K 1.2 M
jam 12.7K 416 68.5K 15.8K 3.5M 378.4K 25.8M 1.7M 609.5 K 7.6 M
gui 1.1K 409 62.7K 24.8K 915K 314K 6.5M 2M 619.3 K 2.1 M
nia 2K 408 2K 25K 1.7M 476.5K 11.3M 3.1M 1.0 M 3.9 M
mas 15.2K 405 216.8K 17.6K 6.2M 390.1K 42.1M 3M 927.5 K 13.4 M
bzj 983 404 33.6K 26.4K 824.3K 565K 4.5M 2.9M 981.2 K 1.4 M
mkn 956 402 33.1K 25.4K 584.2K 456.9K 3.4M 2.6M 734.8 K 1.0 M
lhu 46K 377 975K 15.7K 29.1M 441.2K 208.6M 2.5M 623.0 K 38.8 M
ctu 690 366 35.5K 20.6K 646.7K 352.8K 3.6M 2M 614.9 K 1.2 M
kg 4.7K 365 85.5K 21.7K 2.5M 406.7K 16.6M 2.6M 905.4 K 5.7 M
inb 387 343 17.3K 17K 202.8K 197K 2M 1.9M 535.2 K 555.6 K
guh 1.9K 331 104.9K 28.4K 1.5M 328.4K 11.2M 3M 789.5 K 3.5 M
rn 8.2K 323 8.2K 11.1K 4.5M 179K 33.2M 1.3M 449.9 K 11.8 M
bus 467 322 21.4K 12.1K 418.4K 219.2K 2.1M 1.1M 428.8 K 830.9 K
mfe 7.5K 320 198.8K 18.2K 4.6M 374.8K 26.9M 2.1M 716.4 K 10.1 M
sda 1.6K 317 43.2K 6.2K 2.5M 218.3K 15.8M 1.6M 529.0 K 4.7 M
bi 71.9K 311 308.5K 13.6K 19.4M 359.4K 132.4M 1.9M 546.9 K 42.6 M
cr_Latn 19K 303 170K 8.9K 19K 303 81.8K 1K 590.4 K 15.0 M
gor 1.7K 303 53.3K 6.5K 1.4M 227.1K 9.4M 1.7M 494.0 K 3.1 M
jac 8.2K 303 61.6K 11.9K 1.8M 271K 15.7M 1.7M 530.3 K 7.3 M
chr 964 301 33.8K 7.5K 629.9K 172.3K 4.7M 1M 564.1 K 2.1 M
mh 4.6K 296 235.1K 13K 3.6M 393.5K 24.9M 2.2M 778.4 K 8.4 M
mni 1.2K 290 38.1K 13.2K 841.3K 245.5K 6.4M 1.8M 866.6 K 3.0 M
wal 2.6K 286 128K 14K 2M 203.4K 17M 1.7M 525.7 K 5.1 M
teo 2.8K 274 131.5K 13.7K 2.3M 221.4K 15.3M 1.6M 564.9 K 5.3 M
gub 31.7K 271 160.4K 25K 4.7M 286.2K 44.7M 1.6M 431.3 K 23.1 M
qvi 1.2K 266 48.4K 19.3K 720.4K 248.9K 6.5M 2.3M 641.2 K 1.9 M
tdx 1.7K 262 26.3K 13.2K 1M 238.5K 7M 1.6M 503.6 K 2.1 M
rki 331 251 331 7.8K 119.7K 113.7K 1.6M 1.5M 751.3 K 781.8 K
djk 560 246 30.9K 24.4K 669.5K 455.6K 3.7M 2.2M 644.3 K 1.0 M
nr 10.7K 246 10.7K 11.3K 5.3M 162.5K 49M 1.5M 519.7 K 17.8 M
zne 1.3K 239 61.9K 21.3K 1.4M 504.6K 8.2M 2.8M 882.3 K 2.8 M
izz 423 237 21.7K 14.5K 382.8K 194.5K 2.1M 1.1M 382.2 K 789.9 K
noa 902 234 902 11.5K 821.1K 243.9K 5.2M 1.6M 534.3 K 1.7 M
bqc 275 228 9.8K 8.2K 193K 151.7K 997K 788.4K 317.0 K 408.1 K
srm 847 227 847 17.3K 1.2M 445.3K 6.3M 2M 613.4 K 1.7 M
niq 26.7K 226 26.7K 4.2K 9.9M 103.4K 72.1M 716.2K 239.1 K 20.9 M
bas 4.2K 216 105.2K 14.9K 4.3M 362.8K 25.7M 1.7M 600.7 K 7.6 M
dwr 452 215 22.1K 11.1K 269.4K 139.5K 2.2M 1.2M 375.4 K 747.6 K
guc 537 214 22.9K 12.5K 422.4K 218.1K 3.4M 1.8M 540.1 K 1.1 M
jvn 1K 213 36.2K 7.8K 790.5K 185.6K 5.3M 1.2M 357.2 K 1.7 M
hvn 737 200 33.9K 7K 779.7K 239.4K 4.3M 1.2M 378.5 K 1.4 M
sxn 587 197 587 9.9K 494K 220.6K 3.4M 1.5M 507.1 K 1.2 M
koi 20.7K 196 153.9K 5K 2.2M 89.9K 17.1M 664.5K 323.0 K 7.1 M
alz 2.2K 195 59.3K 12.2K 1.3M 246.9K 7.9M 1.4M 488.1 K 2.9 M
nyu 1.2K 195 1.2K 11K 988.7K 210.5K 7.7M 1.6M 492.6 K 2.2 M
bn_Latn 98.7K 191 1.3M 12K 98.7K 191 458K 730 314.7 K 81.0 M
suz 226 186 226 11.3K 169.6K 140.5K 1M 855.2K 339.5 K 429.6 K
pau 1.7K 185 1.7K 13.1K 2M 394.6K 12.4M 2M 600.1 K 3.2 M
nij 1K 183 1K 9.2K 741.6K 186.1K 4.7M 1.2M 389.6 K 1.6 M
sat_Latn 39K 183 39K 5.5K 39K 183 183.8K 601 276.1 K 39.2 M
gu_Latn 58.2K 179 688.4K 5.4K 58.2K 179 260.8K 673 241.0 K 47.9 M
msm 520 177 520 8.6K 410.8K 190.5K 2.5M 1.1M 339.7 K 789.8 K
maz 585 170 21.3K 8.2K 452.9K 174K 2.9M 951.7K 304.7 K 971.4 K
qxr 2.6K 153 40.8K 6.4K 761.5K 75.4K 6.6M 724K 186.4 K 1.9 M
shp 874 150 22.4K 3.7K 534.1K 96.8K 3.8M 710.4K 216.9 K 1.2 M
hne 3K 146 118.4K 4.3K 2.3M 139.3K 12M 697K 379.3 K 6.5 M
ktu 3.3K 144 115.5K 7.8K 3.2M 196.9K 18.5M 1.1M 300.1 K 5.4 M
laj 6.5K 144 61K 6.4K 2.4M 140.1K 15.8M 730.5K 233.5 K 4.6 M
pis 1.1K 139 62K 7.2K 1.3M 136.8K 7.7M 764K 212.7 K 2.2 M
mag 631 138 62.6K 22.1K 2.1M 544.2K 10.7M 2.6M 1.4 M 5.4 M
gbm 2.5K 137 50.8K 3.8K 1.7M 99.7K 9.1M 499.6K 282.4 K 4.5 M
tzj 471 136 11.1K 7.3K 299.9K 150.8K 1.9M 884.2K 272.0 K 663.9 K
oj 2.5K 135 2.5K 1.6K 1.2M 35.9K 9.6M 337.1K 117.6 K 3.4 M
ndc_ZW 2.2K 132 2.2K 8.7K 2.2K 132 9.1K 523 343.1 K 2.2 M
tks 63.7K 127 63.7K 6.8K 17.1M 41.5K 88.9M 260.8K 39.5 K 33.0 M
awa 5.8K 126 100.1K 8.4K 2.2M 98.7K 11.1M 475K 226.6 K 5.8 M
gvl 37.9K 126 213K 6.9K 21.1M 161.1K 141M 789.2K 257.8 K 31.7 M
knj 229 126 10.1K 9.2K 202.6K 171.8K 1.1M 855K 253.1 K 345.4 K
spp 733 123 733 5.8K 902.7K 141.8K 4.4M 682.5K 217.8 K 1.4 M
mqy 69.3K 119 309K 2.5K 12.1M 88.6K 78.9M 506.5K 170.4 K 16.3 M
tca 410 117 20K 7.3K 283K 121.5K 2.3M 786K 226.2 K 781.2 K
cce 847 116 23.2K 11K 539.3K 227.2K 3.3M 1.3M 393.8 K 1.1 M
skr 3.8K 107 279.3K 17.1K 6.2M 324K 32.2M 1.7M 768.5 K 15.4 M
kmz_Latn 24K 106 361K 2.4K 24K 106 108.6K 401 231.8 K 16.7 M
dje 913 100 40.2K 3.7K 816.3K 97.5K 4.7M 480.7K 161.2 K 1.5 M
gof 2.8K 97 33.8K 5.5K 703K 68.8K 5.5M 506K 159.1 K 1.7 M
agr 465 93 16.1K 3.6K 295.4K 67.2K 2.3M 554.5K 177.0 K 760.1 K
qvz 534 88 6.8K 3.5K 145.5K 50.5K 1.2M 438.3K 124.2 K 382.7 K
adh 2.6K 87 107.2K 1K 2.4M 42.1K 14.5M 254.9K 84.6 K 5.0 M
quf 522 86 8.4K 5.2K 155.7K 61.8K 1.5M 609K 173.7 K 542.8 K
kjg 113 84 3K 2.9K 67.6K 67K 408.5K 399K 159.2 K 167.7 K
tsc 12.6K 82 12.6K 4K 3.5M 93.1K 23.4M 521.3K 161.9 K 7.0 M
ber 2.7K 79 12.6K 1.2K 1.1M 46.4K 6.4M 265.9K 141.5 K 3.0 M
ify 611 79 19.8K 2.8K 422.7K 56.2K 2.6M 334K 109.5 K 913.1 K
cbk 10.1K 78 43.8K 2K 1.7M 64.3K 10.3M 339.3K 93.4 K 3.4 M
quy 588 78 28.1K 2.7K 423.3K 37.3K 4.5M 368.2K 114.5 K 1.2 M
ahk 244 77 6.2K 4.1K 264K 124.8K 1.3M 715.5K 182.8 K 359.7 K
cac 212 77 3.4K 1.8K 125.7K 54.1K 978.7K 319.8K 95.8 K 280.3 K
akb 1K 71 21.3K 408 870.9K 54.5K 5.2M 337.8K 93.7 K 1.6 M
nut 29K 67 29K 1.5K 4.8M 39.8K 23.5M 184.1K 36.4 K 8.3 M
ffm 1.8K 65 30.1K 2K 745.6K 39.1K 4.6M 236.1K 83.8 K 1.8 M
taj 146 65 21.6K 14.3K 309.7K 203K 2.3M 1.4M 503.0 K 872.7 K
ms_Arab 698 63 698 320 698 63 2.9K 239 64.7 K 1016.0 K
brx 322 62 5.3K 2.4K 144.2K 41K 1.1M 304.4K 146.6 K 515.7 K
ann 464 56 5K 1.6K 116.4K 35.9K 760.9K 215.1K 74.9 K 295.2 K
qup 169 53 4.3K 2.5K 77.5K 31.3K 763.8K 297.8K 74.7 K 207.3 K
ms_Arab_BN 2.6K 46 2.6K 374 2.6K 46 10.5K 171 50.0 K 5.1 M
miq 236 45 6.4K 3.5K 183.7K 80.2K 1.2M 485.6K 157.6 K 384.1 K
msb 811 41 811 1K 705.9K 28.8K 4.4M 167.5K 53.3 K 1.7 M
bim 410 40 31.1K 6.3K 669.8K 167.4K 3.2M 793.4K 252.7 K 1.1 M
raj 1.8K 40 1.8K 5.7K 1.3M 81.1K 7.1M 405K 226.2 K 3.9 M
kwi 382 37 16.9K 2.2K 253.8K 23.4K 1.8M 172.8K 47.6 K 536.2 K
tll 200 37 200 2.7K 304.2K 62.2K 2.2M 409.8K 132.3 K 664.5 K
trp 12.8K 36 12.8K 1.7K 4.1M 39K 29.9M 257.3K 87.5 K 10.2 M
smt 1.4K 34 1.4K 703 1M 36.5K 6.8M 245.4K 87.9 K 2.5 M
mrw 11.3K 29 11.3K 1K 4.2M 45.7K 27.8M 257.2K 81.3 K 8.8 M
dln 236 28 5.2K 969 150.8K 21.5K 860.5K 118.3K 36.8 K 280.3 K
qvc 3.4K 27 14.6K 2.2K 495.7K 25.7K 5M 233.7K 65.3 K 2.6 M
doi 1.7K 26 21.8K 975 568.7K 25.5K 3.2M 135.3K 66.7 K 1.6 M
ff 13.6K 26 150K 5K 3.4M 46.5K 22.8M 277.6K 78.8 K 8.5 M

Citation Information 

@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}
}