Glot500 Corpus

A dataset of natural language data collected by putting together more than 150 existing mono-lingual and multilingual datasets together and crawling known multilingual websites. The focus of this dataset is on 500 extremely low-resource languages. (More Languages still to be uploaded here)

This dataset is used to train the Glot500 model.

• Homepage: homepage
• Repository: github
• Paper: acl, arxiv

This dataset has the identical data format as the Taxi1500 Raw Data dataset, so that both datasets can be used in parallel seamlessly.

Parts of the original Glot500 dataset cannot be published publicly. Please fill out [thi form]{https://docs.google.com/forms/d/1FHto_4wWYvEF3lz7DDo3P8wQqfS3WhpYfAu5vM95-qU/viewform?edit_requested=true} to get access to these parts.

Usage

Replace nbl_Latn with your specific language.

from datasets import load_dataset
dataset = load_dataset('cis-lmu/Glot500', 'nbl_Latn', split='train')
print(dataset['train'][0]) # First row of nbl_Latn

Click to show supported languages:

ton_Latn
nld_Latn
tzo_Latn
leh_Latn
cuk_Latn
ibg_Latn
uzb_Cyrl
jav_Latn
rap_Latn
zpa_Latn
bak_Cyrl
por_Latn
quy_Latn
ast_Latn
cos_Latn
fon_Latn
sna_Latn
dzo_Tibt
nob_Latn
nch_Latn
ish_Latn
che_Cyrl
ext_Latn
ldi_Latn
dtp_Latn
yue_Hani
kbd_Cyrl
mar_Deva
ron_Latn
acr_Latn
afb_Arab
sqi_Latn
eng_Latn
ksd_Latn
rus_Cyrl
bcl_Latn
ksh_Latn
hin_Latn
myv_Cyrl
kjh_Cyrl
sah_Cyrl
gkp_Latn
naq_Latn
tdt_Latn
rmn_Cyrl
kac_Latn
cak_Latn
kir_Cyrl
mps_Latn
yid_Hebr
dhv_Latn
srn_Latn
div_Thaa
mkd_Cyrl
idu_Latn
bre_Latn
bas_Latn
ven_Latn
pxm_Latn
wuu_Hani
mwl_Latn
miq_Latn
kss_Latn
wes_Latn
slv_Latn
hrv_Latn
hmo_Latn
som_Latn
bod_Tibt
pls_Latn
ile_Latn
luo_Latn
pus_Arab
fao_Latn
fas_Arab
swa_Latn
ifb_Latn
ary_Arab
tbz_Latn
hus_Latn
ote_Latn
ilo_Latn
ctd_Latn
abk_Cyrl
bqc_Latn
hil_Latn
pon_Latn
zul_Latn
als_Latn
pes_Arab
bpy_Beng
bos_Latn
sot_Latn
lin_Latn
tuk_Cyrl
gla_Latn
wln_Latn
apc_Arab
hin_Deva
hye_Armn
tir_Ethi
pap_Latn
gcf_Latn
cjk_Latn
pcd_Latn
tur_Latn
kon_Latn
mwn_Latn
izz_Latn
xho_Latn
lam_Latn
guc_Latn
aka_Latn
kea_Latn
sme_Latn
fat_Latn
csb_Latn
bak_Latn
djk_Latn
xav_Latn
oci_Latn
acm_Arab
rmy_Cyrl
bim_Latn
mck_Latn
krc_Cyrl
cym_Latn
lus_Latn
ncx_Latn
ngu_Latn
yom_Latn
tam_Taml
ajp_Arab
epo_Latn
fra_Latn
ita_Latn
seh_Latn
sxn_Latn
pdt_Latn
hbs_Latn
uzn_Cyrl
bhw_Latn
ksw_Mymr
pms_Latn
zlm_Latn
ami_Latn
qub_Latn
twx_Latn
tsz_Latn
kaa_Cyrl
toj_Latn
toh_Latn
kos_Latn
ogo_Latn
kab_Latn
pan_Guru
nan_Latn
aze_Latn
prk_Latn
ara_Arab
meu_Latn
nba_Latn
lvs_Latn
nbl_Latn
loz_Latn
crh_Latn
bci_Latn
kbp_Latn
tgl_Latn
kmb_Latn
hun_Latn
nzi_Latn
yao_Latn
arn_Latn
hyw_Cyrl
vmw_Latn
jbo_Latn
mzn_Arab
lzh_Hani
heb_Hebr
cce_Latn
bjn_Latn
gug_Latn
yor_Latn
ban_Latn
tlh_Latn
chv_Cyrl
sin_Sinh
ind_Latn
dua_Latn
sid_Latn
amh_Ethi
zea_Latn
kpg_Latn
crh_Cyrl
nyu_Latn
dln_Latn
ibo_Latn
tih_Latn
msa_Latn
nap_Latn
mgr_Latn
bik_Latn
srp_Cyrl
lao_Laoo
guw_Latn
kom_Cyrl
sop_Latn
nde_Latn
hui_Latn
cfm_Latn
new_Deva
kur_Arab
sco_Latn
nyk_Latn
lun_Latn
suz_Deva
wal_Latn
asm_Beng
rar_Latn
san_Deva
kaz_Cyrl
tog_Latn
iba_Latn
tuk_Latn
nso_Latn
run_Latn
ctu_Latn
bam_Latn
fin_Latn
gor_Latn
kmr_Latn
ben_Beng
pag_Latn
niu_Latn
xmf_Geor
ekk_Latn
tsc_Latn
lmo_Latn
mhr_Cyrl
plt_Latn
qvi_Latn
roh_Latn
oke_Latn
mah_Latn
tok_Latn
mgh_Latn
eml_Latn
urh_Latn
pnb_Arab
yua_Latn
nav_Latn
zne_Latn
bin_Latn
cat_Latn
gym_Latn
sat_Olck
snd_Arab
isl_Latn
rmn_Grek
bba_Latn
kal_Latn
aoj_Latn
qug_Latn
zai_Latn
guj_Gujr
min_Latn
tob_Latn
grc_Grek
hmn_Latn
ido_Latn
khm_Khmr
ikk_Latn
iku_Cans
tat_Latn
bel_Cyrl
dyu_Latn
que_Latn
efi_Latn
quw_Latn
nyn_Latn
wol_Latn
hne_Deva
zho_Hani
swh_Latn
bum_Latn
kua_Latn
ncj_Latn
ewe_Latn
hat_Latn
ina_Latn
mfe_Latn
ahk_Latn
srm_Latn
lug_Latn
ach_Latn
rmy_Latn
tpm_Latn
smo_Latn
mos_Latn
srd_Latn
srp_Latn
azb_Arab
ori_Orya
mzh_Latn
kur_Latn
phm_Latn
kwn_Latn
crs_Latn
ada_Latn
ttj_Latn
hif_Latn
tzh_Latn
tdx_Latn
bbc_Latn
cnh_Latn
pcm_Latn
tso_Latn
nor_Latn
bsb_Latn
kqn_Latn
gaa_Latn
ukr_Cyrl
lav_Latn
nep_Deva
kmr_Cyrl
ige_Latn
pis_Latn
lhu_Latn
nya_Latn
tiv_Latn
mny_Latn
kri_Latn
nyy_Latn
poh_Latn
nnb_Latn
grn_Latn
mco_Latn
ory_Orya
ful_Latn
diq_Latn
sag_Latn
tel_Telu
afr_Latn
haw_Latn
umb_Latn
hsb_Latn
fij_Latn
hbs_Cyrl
san_Latn
vls_Latn
zsm_Latn
lij_Latn
quc_Latn
mam_Latn
tuc_Latn
dan_Latn
rue_Cyrl
ace_Latn
bem_Latn
kam_Latn
ndo_Latn
mbb_Latn
mrw_Latn
ajg_Latn
oss_Cyrl
her_Latn
lit_Latn
frr_Latn
yap_Latn
bzj_Latn
gom_Latn
swe_Latn
lfn_Latn
cmn_Hani
mon_Cyrl
vep_Latn
ixl_Latn
gil_Latn
mau_Latn
aym_Latn
gom_Deva
fur_Latn
cgg_Latn
chw_Latn
kin_Latn
alz_Latn
ndc_Latn
gcr_Latn
rmn_Latn
sgs_Latn
bih_Deva
skg_Latn
bts_Latn
vie_Latn
tha_Thai
tcf_Latn
pau_Latn
est_Latn
lue_Latn
rug_Latn
gur_Latn
kik_Latn
mri_Latn
ber_Latn
ssw_Latn
cab_Latn
quz_Latn
arb_Arab
mai_Deva
tat_Cyrl
mya_Mymr
alt_Cyrl
nno_Latn
nse_Latn
hrx_Latn
hau_Latn
koo_Latn
gsw_Latn
pam_Latn
sun_Latn
lat_Latn
bis_Latn
btx_Latn
udm_Cyrl
xmv_Latn
tca_Latn
uig_Arab
glg_Latn
tah_Latn
llb_Latn
ckb_Arab
gle_Latn
lim_Latn
slk_Latn
nds_Latn
kor_Hang
uzb_Latn
gkn_Latn
pfl_Latn
azj_Latn
glv_Latn
jam_Latn
kat_Geor
abn_Latn
fry_Latn
kat_Latn
twi_Latn
eus_Latn
toi_Latn
mlg_Latn
ifa_Latn
tyv_Cyrl
arz_Arab
chk_Latn
vol_Latn
kek_Latn
teo_Latn
ell_Grek
kan_Knda
rng_Latn
tpi_Latn
mdy_Ethi
lua_Latn
mad_Latn
top_Latn
scn_Latn
ngl_Latn
mal_Mlym
szl_Latn
orm_Latn
nia_Latn
urd_Arab
mxv_Latn
cbk_Latn

License

We don't own any part of the data. The original source of each sentence of the data is indicated in dataset field.

To see the copyright license of the original datasets visit here.

We license the actual packaging, the metadata and the annotations of these data under the cc0-1.0.

If you are a website/dataset owner and do not want your data to be included in this corpra, please send us an email at glot500@cis.lmu.de.

Ethical Considerations

1. Biases: The text corpus may reflect the perspectives, opinions, or demographics of its sources or creators. It is important for users to critically evaluate the text in context especially for news sources and social medias.

2. Representativeness: While we have aimed for diversity and inclusivity, the text corpus may not fully represent all native speakers. Users should be mindful of any potential underrepresentation.

3. Ethics: We acknowledge that the collection and use of text data can have ethical implications. We have strived to handle the data responsibly, but we encourage users to consider the broader ethical implications of their own research or applications.

Citation

If you use any part of this code and data in your research, please cite it using the following BibTeX entry.

@inproceedings{imanigooghari-etal-2023-glot500,
    title = "Glot500: Scaling Multilingual Corpora and Language Models to 500 Languages",
    author = {ImaniGooghari, Ayyoob  and
      Lin, Peiqin  and
      Kargaran, Amir Hossein  and
      Severini, Silvia  and
      Jalili Sabet, Masoud  and
      Kassner, Nora  and
      Ma, Chunlan  and
      Schmid, Helmut  and
      Martins, Andr{\'e}  and
      Yvon, Fran{\c{c}}ois  and
      Sch{\"u}tze, Hinrich},
    editor = "Rogers, Anna  and
      Boyd-Graber, Jordan  and
      Okazaki, Naoaki",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.61",
    doi = "10.18653/v1/2023.acl-long.61",
    pages = "1082--1117",
    abstract = "The NLP community has mainly focused on scaling Large Language Models (LLMs) vertically, i.e., making them better for about 100 languages. We instead scale LLMs horizontally: we create, through continued pretraining, Glot500-m, an LLM that covers 511 predominantly low-resource languages. An important part of this effort is to collect and clean Glot500-c, a corpus that covers these 511 languages and allows us to train Glot500-m. We evaluate Glot500-m on five diverse tasks across these languages. We observe large improvements for both high-resource and low-resource languages compared to an XLM-R baseline. Our analysis shows that no single factor explains the quality of multilingual LLM representations. Rather, a combination of factors determines quality including corpus size, script, {``}help{''} from related languages and the total capacity of the model. Our work addresses an important goal of NLP research: we should notlimit NLP to a small fraction of the world{'}s languages and instead strive to support as many languages as possible to bring the benefits of NLP technology to all languages and cultures. Code, data and models are available at \url{https://github.com/cisnlp/Glot500}.",
}