Large filtered sentence-level corpus (web-mined pairs filtered by alignment score, language and length ratio; encyclopedic text; short everyday sentences; talks; software UI strings) plus a document-level subset of step-by-step reasoning texts.

Intended as the first (pre-training) stage; fine-tune afterwards on the quality corpus.

column	description
`uz`	Uzbek text
`other`	English text
`other_lang`	`en`
`bt`	always `false` here
`level`	`sentence` or `document`

Deduplicated; pairs overlapping the quality validation/test splits removed. Uzbek apostrophes: o‘ g‘ (U+2018) and ’ (U+2019) for the tutuq sign.

Attribution and licenses

This corpus is a filtered compilation of the following datasets; each part keeps its original license. Parts under CC-BY-SA must be shared under the same license.

part	license
NLLB (Meta AI mined bitext, via OPUS) — Schwenk et al. 2021; NLLB Team 2022	ODC-By 1.0
Wikimedia en-uz (via OPUS)	CC-BY-SA 4.0
UzLPC	CC-BY-SA 4.0
KDE4, GNOME (via OPUS)	licenses of the original software
translatewiki.net (via OPUS)	CC BY 3.0
tldr-pages (via OPUS)	CC BY 4.0
Tatoeba (via OPUS)	CC BY 2.0 FR
reasoning subset: glaiveai/reasoning-v1-20m, Uzbek side machine-translated	Apache-2.0

OPUS: J. Tiedemann, 2012, Parallel Data, Tools and Interfaces in OPUS, LREC 2012.

Downloads last month: 20

Total file size:

1.16 GB