Datasets:

failed09
/

bashkir-frequency-index

license: cc-by-4.0
language:
  - ba
pretty_name: Bashkir Frequency Index (83M tokens, RU-filtered press corpus)
task_categories:
  - text-classification
configs:
  - config_name: public
    data_files:
      - split: train
        path: frequency_public.parquet
  - config_name: full
    data_files:
      - split: train
        path: frequency_full.parquet
tags:
  - bashkir
  - frequency
  - wordlist
  - lexicon
  - low-resource
  - cyrillic
  - corpus-statistics

Bashkir Frequency Index v6.1

Word-frequency index for the Bashkir language, built from 83.27 million tokens of cleaned monolingual text (newspapers, magazines, encyclopedia, Wikipedia, literature, speech transcripts; Russian-language admixture removed by sentence-level filtering).

Configurations

config	rows	cutoff	use case
`public` (recommended)	521,415	count ≥ 3	applications, spellcheckers, T9, research
`full`	1,290,098	count ≥ 1	morphology studies, OCR noise analysis

Additionally: top_coverage_50.json — the minimal set of 1,634 word forms covering 50% of all tokens (with cumulative coverage), handy as a "learn these first" vocabulary list.

Schema

column	type	description
`rank`	int64	position by descending count
`word`	string	word form, lowercase
`count`	int64	absolute frequency (tokens)
`ipm`	float	instances per million tokens
`doc_freq`	int32	number of sentences containing the form

Sample

rank	word	count	ipm	doc_freq
1	һәм	1,086,463	13,046.8	775,593
23	башҡорт	160,963	1,932.9	114,353
126	йылға	59,878	719.0	55,988
390	тел	23,031	276.6	19,225
439	һуғыш	20,963	251.7	17,315

Coverage

tokens covered	forms needed
25%	188
50%	1,634
80%	17,559
90%	53,852
95%	131,559

Method

Tokenization: regex lowercase extraction over the Cyrillic alphabet incl. the 9 Bashkir-specific letters (Ә Ғ Ҙ Ҡ Ң Ө Ҫ Ү Һ).
Corpus base: monolingual master v2 — 5,481,554 sentences; Russian sentences (~11.6% of tokens in the previous version) removed with a function-word filter (ru_filter_v1).
Cleanups vs v5.0: glued-word forms >24 chars dropped (490), 1-char forms dropped except ә / у / а, 34,922 rare forms merged into their Damerau-Levenshtein-1 neighbours (conservative: ratio ≥ 50, length ≥ 5).

v6.1 changelog

Removed 65 Russian function-word forms that leak inside Bashkir sentences (quotes, names): на, не, по, из, это... — 648,304 tokens. The list is homograph-checked against Bashkir (kept: был = "this", ни = "what", ти/тип = quotative, да/ла/лә/ҙа = "also" particle, без = "we").
Removed single-letter forms а, у (Russian-dominant usage); kept ә (the Bashkir conjunction "and").
Baseline cleanups identical to v6.0 (glued forms, typo merge).

Known caveats

The index is wordform-level (not lemmatized): agglutination produces many forms of one lemma.
The long tail (count < 3) contains OCR/source typos and glued artifacts.
Proper names, especially Russian patronymics, appear in the tail.

Loading

from datasets import load_dataset

freq = load_dataset("failed09/bashkir-frequency-index", "public")["train"]

# top-20 verbs starting with "бар"
[f for f in freq.filter(lambda r: r["word"].startswith("бар"))][:20]

Citation

@dataset{failed09_bashkir_frequency_index_2026,
  title  = {Bashkir Frequency Index v6.1},
  author = {failed09},
  year   = {2026},
  publisher = {Hugging Face},
  url    = {https://huggingface.co/datasets/failed09/bashkir-frequency-index}
}