parchiev/ingush-russian

Ingush–Russian Parallel Literary Corpus

A sentence-aligned Ingush (ГӀалгӀай мотт, ISO 639-3 inh) – Russian parallel corpus built from 7 published literary translations. Both sides are human translations, so the data can be used for training and evaluating machine translation.

Source and attribution. This corpus was compiled and aligned by the Ghalghay project («Словарный свод ингушского языка», by DD):

• project site: https://ghalghay.vercel.app/
• about the project: https://ghalghay.vercel.app/about.html
• the parallel corpus and its source list: https://ghalghay.vercel.app/parall.html, https://ghalghay.vercel.app/parall_about.html

Please credit the Ghalghay project (and the translators listed below) whenever you use this dataset.

Ingush is a Nakh language of the North Caucasus, spoken by several hundred thousand people and classified as vulnerable by UNESCO. This dataset is part of an effort to build open, clean Ingush resources for language technology and language preservation.

config	split	rows	works
default	train	4392	Gulliver among the Lilliputians (children's retelling), Mozart and Salieri, The Captain's Daughter, The Tale of Igor's Campaign, The Tales of Belkin: The Squire's Daughter
default	validation	410	Rikki-Tikki-Tavi
default	test	564	Mumu
raw	train	5505	all rows, including excluded ones, with raw markup and the edit log

from datasets import load_dataset

ds = load_dataset("parchiev/ingush-russian")              # cleaned train / validation / test
raw = load_dataset("parchiev/ingush-russian", "raw")      # every original row + provenance (single split)

Languages and orthography

• Ingush in Cyrillic script. The letter palochka is always the invariant Ӏ (U+04C0), including inside lowercase words (гӀалгӀай). Look-alikes (І U+0406, Latin I, digit 1, ӏ U+04CF) were replaced. Note: Python's str.lower() turns Ӏ into ӏ (U+04CF); map it back if you lowercase.
• Russian: literary Russian (the original texts or published Russian translations).
• Unicode NFC; no compatibility (NFKC) folding.

Dataset structure

Example (default, validation):

{
  "id": 21,
  "gh": "Цо бӀаргаш хьа а дийлар, Ӏиршинг а даьлар цунна.",
  "ru": "Он открыл глаза и чихнул.",
  "work_id": "КиплРикк",
  "work_title": "Rikki-Tikki-Tavi",
  "author": "Rudyard Kipling",
  "translator_inh": "Т. И. Мальсагов",
  "year_inh": "1939",
  "ru_version": "Russian translation by К. Чуковский",
  "verse_no": null,
  "is_verse": false,
  "gh_notes": null,
  "ru_notes": null,
  "flags": []
}

field	type	description
id	int	stable row id = row index of the original corpus (0–5504); identical to the row order of v1
gh	string	Ingush sentence/segment, normalized (column name kept from v1 for compatibility)
ru	string	Russian sentence/segment, normalized
work_id	string	work identifier from the source corpus
work_title	string	English title of the work
author	string	author of the original work
translator_inh	string?	translator into Ingush, when known
year_inh	string?	year the Ingush translation was published
ru_version	string	what the Russian side is (original / published translation)
verse_no	string?	verse fragment number (The Tale of Igor's Campaign, Jakobson's division)
is_verse	bool	the segment is poetry; line breaks were joined with spaces
gh_notes, ru_notes	string? (JSON list)	translator/editor footnotes that were removed from the text
flags	list[string]	length_outlier (free translation, check before use) or dup_of:<id> (exact duplicate pair)

The raw config adds split, gh_raw / ru_raw (the source markup), hf_v1_gh / hf_v1_ru (the text as published in v1 of this dataset) and edits (a JSON log of every normalization change: offset, from, to, rule).

Splits

Splits hold out whole works, so that test sentences never share context, characters or translator style with the training data:

• test: Turgenev, Mumu (Ingush translation, 1939)
• validation: Kipling, Rikki-Tikki-Tavi (translated by Т. И. Мальсагов, 1939)
• train: the other 5 works

Source data

The corpus comes from the Ghalghay project's «Корпус параллельных переводов» (parallel translation corpus). It is part of «Словарный свод ингушского языка» (the Ingush dictionary compendium) by DD:

• Project home: https://ghalghay.vercel.app/ (mirror of ghalghay.github.io)
• About the project and its bibliography: https://ghalghay.vercel.app/about.html
• Parallel corpus viewer: https://ghalghay.vercel.app/parall.html
• Sources of the parallel texts: https://ghalghay.vercel.app/parall_about.html

The Ghalghay project digitized, proofread and sentence-aligned the published translations below. v2 of this dataset was rebuilt from its structured original (src_parall/data.js). v1 was a flattened copy of the same corpus.

work	author	Ingush translation	published	rows (all)	rows (default)
The Captain's Daughter	Alexander Pushkin	Х. Осмиев	1961	2930	2898
Gulliver among the Lilliputians (children's retelling)	Jonathan Swift	—	1939	735	701
Mumu	Ivan Turgenev	—	1939	569	564
The Tales of Belkin: The Squire's Daughter	Alexander Pushkin	С.-Г. Котиев	2005	419	416
Rikki-Tikki-Tavi	Rudyard Kipling	Т. И. Мальсагов	1939	413	410
Mozart and Salieri	Alexander Pushkin	С. Чахкиев	2004	220	159
The Tale of Igor's Campaign	anonymous (12th c.)	С. Чахкиев	2004	219	218

The same material is also available in other collections (for example lingtrain/ingush-russian, and the PaydaDosh corpus). Deduplicate before combining.

Processing (v2)

1. Rebuilt from the source markup instead of flattened text. Paragraph (¶) and verse-line (|) separators became spaces, and verse numbers were moved to verse_no. Footnote reference marks (*, 5, …) were removed, and footnote texts were moved to gh_notes / ru_notes.
2. Normalization (see the edit log in raw): SPC-002 × 112, PAL-001 × 74, PUN-008 × 8, UNI-010 × 7, STD-001 × 6, SPC-008 × 2, SPC-004 × 2.
3. Rows excluded from default but kept in raw, with the reason in flags: heading × 62, speaker_label × 42, missing_side × 25, numeric_only × 22, untranslated × 10. missing_side marks passages the Ingush translator left out (shown as <…> in the source).
4. No Ingush words were hand-corrected. Spelling is as published, including historical (1939) spellings.

A before/after audit is in the changelog.

Known limitations

• Literary register only (19th-century prose, verse, a medieval epic, children's literature). Little conversational or modern vocabulary.
• Translations are sometimes free, and some 1:many sentence splits remain (length_outlier flags 12 rows).
• 1939 texts use older spelling conventions. Some OCR errors may remain.
• Small size: best used for fine-tuning and evaluation, together with other Ingush data.

Rights and licensing

The texts are published literary translations. Copyright in the Ingush and Russian texts remains with their authors, translators and publishers. This compilation is shared for non-commercial research, education and Ingush language preservation. It is not an open-content license. If you are a rights holder and want a text removed, please open a discussion on this repository.

The earlier license: mit tag was removed, because it could not apply to third-party literary texts.

Citation

@misc{parchiev_ingush_russian_2026,
  title        = {Ingush--Russian Parallel Literary Corpus (v2)},
  author       = {Parchiev, Rauf},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/parchiev/ingush-russian}},
  note         = {Rebuilt from the Ghalghay project's parallel corpus, https://ghalghay.vercel.app/}
}

@misc{ghalghay_parallel_corpus,
  title        = {Словарный свод ингушского языка: Корпус параллельных переводов},
  author       = {Ghalghay project (DD)},
  howpublished = {\url{https://ghalghay.vercel.app/}},
  note         = {About: https://ghalghay.vercel.app/about.html}
}

Acknowledgements

• The Ghalghay project (https://ghalghay.vercel.app/, https://ghalghay.vercel.app/about.html) digitized and aligned this corpus. All credit for the alignment goes to them.
• The translators: Т. И. Мальсагов, Х. Осмиев, С. Чахкиев, С.-Г. Котиев, and the unnamed 1939 translators, together with the authors of the original works and of the Russian translations (К. Чуковский, Д. С. Лихачёв).

Changelog

• v2 (2026-09): rebuilt from the structured source. Normalization, footnotes and verse numbers separated, non-parallel rows excluded, work-level train/validation/test splits, metadata columns, raw config, and an English dataset card. gh/ru keep their names. data/train-00000-of-00001.parquet now holds only the train split. v1 remains available as the git tag v1 (load_dataset("parchiev/ingush-russian", revision="v1"), commit 76e84ae).
• v1 (2024-05): 5,505 rows, gh/ru, single train split.

Audit (v1 → v2 default):

check	v1	v2
rows with palochka look-alikes (I, 1, l, ӏ, І) in Ingush	79	0
rows with mixed Latin/Cyrillic words in Ingush	13	0
rows with HTML entities	29	0
rows with single * footnote markers (Pushkin's *** name ellipses are kept)	24	0
rows with | verse separators	562	0
rows with leading/trailing/double spaces	123	0