vpermilp/lid-udm

spellman LID corpus — udm

83,867 rows / 21.1M chars of udm-language text with provenance, built for training and evaluating the spellman language detector. Rows are single-line texts: news articles and library works (paragraph/sentence-group sized) plus Telegram/VK community posts and comments (wild register — emoji, hashtags, code-switching artefacts, folk spellings kept on purpose).

Row length: median 130 chars (p10 30, p90 602).

Configs

config	contents	rows
moderated (default)	news · library · wiki — edited, published sources	0
wild	tg-post · tg-comment · vk-post · vk-comment — Telegram and VK community posts and comments	83,867
all	union of both	83,867

from datasets import load_dataset
moderated = load_dataset("vpermilp/lid-udm")            # default config
wild = load_dataset("vpermilp/lid-udm", "wild")

Columns

column	meaning
text	the row (single line, whitespace-flattened)
lang	ISO 639-3 (udm)
source	site store or Telegram channel handle
register	news · library · wiki · tg-post · tg-comment · vk-post · vk-comment
date	publication date when the source exposed one (YYYY-MM-DD)
id	source URL (crawl) or message/post id (Telegram, VK)

Registers

register	rows
vk-post	82,581
tg-post	1,190
tg-comment	96

Sources

source	register	rows
udmurteveryday	vk-post	24,207
myudm	vk-post	18,009
udmdunne	vk-post	15,666
udmkylburchi	vk-post	5,434
shundy	vk-post	3,001
udmurtkenesh	vk-post	2,839
daurtv	vk-post	2,690
vordskem_kyl	vk-post	2,069
zechbur_gazet	vk-post	1,728
kenesh_lit	vk-post	1,706
udmurtnylkyshno	vk-post	1,365
sharakuara	vk-post	1,148
kiziliudm	vk-post	988
invozho_journal	vk-post	955
udmurt_kyrzan	vk-post	776
udmurtchat	tg-post	465
invozho_journal	tg-post	296
uralistica_com	tg-post	197
malylenlyez	tg-post	116
invozho_journal	tg-comment	33
udmurtslow	tg-comment	29
daurmedia	tg-post	27
udmurtkylmy	tg-comment	27
udmurtslow	tg-post	26
udmdunne	tg-post	26
ivorudmurt	tg-post	23
udmurtkylmy	tg-post	14
malylenlyez	tg-comment	5
daurmedia	tg-comment	2

How it was built and cleaned

1. Orthographic gate — every row carries letters that never occur in Russian orthography (e.g. Sakha ө ү ҥ ҕ һ), with a per-language marker-density floor; Russian-function-word share gates code-switched rows, which are split into native sentence runs rather than dropped.
2. Structural cleaning — site chrome and per-channel boilerplate (lines repeated across ≥8 posts), self-promo tails, TV schedules, Telegram thread roots (auto-forwarded post copies) removed.
3. Structural repairs — soft hyphens removed and print-layout splits rejoined (hyphenated / lowercase-continued paragraphs merged in page order); site breadcrumbs, TV schedules, weather tables, chord sheets, sponsor tails and letter-spacing artefacts dropped; Latin h/x homoglyphs inside Cyrillic words normalized (40 rows dropped); long rows must carry marker letters in ≥2 tokens (0).
4. Content-free filter — emoji/URL/number-only rows, Latin-dominant or mojibake rows, HTML-entity leftovers dropped (99).
5. Cross-source dedup on a normalized key (casefold; URLs, handles, hashtags, punctuation, digits stripped) — catches site↔channel cross-posts and reposts (1,359 removed).
6. PII masking — phone numbers and e-mails replaced by <phone> / <email>; no author identities were ever stored.
7. Model judge — rows the current spellman model contradicts at ≥0.9 confidence outside the language's close-pair group are dropped (12).

No train/validation/test split is shipped: the spellman pipeline assigns splits content-addressed (crc32) at mix time.

License

The compilation, cleaning and annotations are released under the MIT license. Row texts are redistributed from the public sources listed above (newspapers, libraries, public Telegram channels and VK communities) for language-identification research; each row carries its source and id so upstream attribution is always recoverable.

Built 2026-08-30T12:29:30+00:00.