Datasets:
Женщина Дагестана — Russian–Dagestani parallel corpus
Sentence pairs mined from the print archive of Женщина Дагестана ("Woman of Dagestan"), a magazine published in Russian and six Cyrillic-script Dagestani languages.
21,615 rows, issues 2016–2026. One row per Russian sentence; each language's rendering of it sits in its own column, empty where that edition does not carry it.
| column | language | sentences |
|---|---|---|
rus |
Russian (pivot) | 21,615 |
lez |
Lezgian | 21,615 |
Three columns and nothing else: the issue, and the sentences.
How it was built
- OCR. The archive is scanned, and the PDFs' own text layer is unusable — clean-looking Cyrillic with the columns silently interleaved — so all 19,550 pages went to a vision model.
- Article matching. The editions are not one layout translated: they carry different article sets in a different order, and in one issue two articles are simply swapped between the Russian and Lezgian editions. Articles are matched bipartitely, on the names and numbers that survive translation, weighted by how rare each is within the issue.
- Sentence alignment. Within a matched article the order does hold, so: Gale & Church on character length, with a bonus for a shared name or number.
- Filtering. Every edition prints the same Russian insert untranslated — recipes, horoscopes,
beauty. 17% of "Lezgian" pages contain no Lezgian at all, and left in they align against their
own Russian originals as the highest-confidence rows in the corpus. They are removed at both
article and sentence level by counting letters Russian does not have (
Ӏ уь аь иь оь гъ къ хъ кь гь). - Judging. An LLM judge scored every pair 0-3 on whether the target sentence really carries the Russian one. Mixed invisibly into its batches were 60 deliberately mismatched pairs: it scored 100% of them 0, and accepted 0% of them at the threshold used here (
>= 2). Rows below the threshold were dropped.
Limitations — read these before training on it
- The editions are adaptations, not translations. Sentences get rewritten, merged and dropped. Every row here has already survived an LLM judge, but some will still correspond only loosely.
- The text is OCR and has not been proof-read by a speaker of any of these languages.
- The palochka is printed as a bare vertical stroke, so the vision model returns Latin
I; it is normalised toӀ(U+04C0). Residual errors are possible. - Coverage is uneven. Only the languages above are aligned; the rest of the archive is OCRed but not yet matched.
Licensing
CC BY-NC 4.0. Use it, share it, build on it, cite it — but not commercially.
Non-commercial is the ceiling, not a preference. The articles are © Женщина Дагестана / Агентство информации и печати Республики Дагестан (https://женщинадагестана.рф/), and this corpus is derived from them. It is released for research and language preservation, in the belief that this qualifies as свободное использование for scientific purposes (ГК РФ ст. 1274). The magazine is distributed free by a state agency and its archive is public — but public is not public domain, and commercial reuse of the text is not the maintainer's to grant. If you are the rightsholder and object, open an issue on this dataset and it will be taken down.
Citation
@misc{dagwoman2026,
title = {Женщина Дагестана: a Russian–Dagestani parallel corpus},
author = {Asvarov, Alidar},
year = {2026},
note = {Mined from the print archive of Женщина Дагестана (© the publisher).
Released under CC BY-NC 4.0.},
url = {https://huggingface.co/datasets/alidar/dagwoman-parallel}
}
- Downloads last month
- 10