failed09/bashkir-ngram-index

Bashkir Word N-gram Index v13.0-af1

Exact within-sentence word n-gram counts for Bashkir: unigrams, bigrams and trigrams for spellchecking, OCR post-processing and lightweight language modelling.

Overview

Exact word n-gram counts derived from a monolingual Bashkir-language dataset. The release provides unigram, bigram and trigram indexes for corpus processing, spellchecking, OCR post-processing, autocomplete and lightweight language-model experiments. The unigrams configuration doubles as standalone word statistics.

At a glance
Task	Word n-gram statistics
Default config	unigrams
Fields	unigrams: rank, word, count, doc_freq, freq_per_million; n-grams: ngram, count
Source	A monolingual Bashkir-language dataset
License	CC BY 4.0

Contents

Files and Configurations

Config / File	Contents	Rows
unigrams	Word forms with rank, count, document frequency and normalized frequency	801,775
bigrams	Two-word sequences with exact counts	10,777,821
trigrams	Three-word sequences with exact counts	14,556,223

Bigram and trigram files contain only combinations with count >= 2, which reduces the impact of one-off noise. All n-grams are counted inside individual sentences; no n-gram crosses a sentence boundary. Counts and hashes are the source of truth in META.json.

Schema

unigrams.parquet:

Column	Type	Description
rank	int64	Rank by descending frequency
word	string	Lowercase Cyrillic word form
count	int64	Absolute occurrence count
doc_freq	int64	Number of sentences containing the form
freq_per_million	double	Normalized frequency per million indexed tokens

bigrams.parquet and trigrams.parquet contain ngram (words separated by spaces) and count.

Examples

Frequent bigrams:

N-gram	Count
шулай уҡ	185,254
бер нисә	93,945
тағы ла	74,170
халыҡ ара	64,669
шул уҡ	58,412

Frequent trigrams:

N-gram	Count
шул уҡ ваҡытта	31,234
бөйөк ватан һуғышы	18,494
тамағынан км өҫтәрәк	18,089
км өҫтәрәк ҡушыла	18,049
ярына тамағынан км	18,018

These are corpus frequencies, not curated phrase lists. Some highly frequent sequences reflect names, formulaic news language, repeated geographic patterns or segmentation artifacts rather than idiomatic phrases.

Method

Tokens outside the retained frequency vocabulary are removed before n-gram counting. Removed tokens can create adjacency inside a sentence; sentence boundaries are never crossed.

upstream cleaned monolingual corpus (13,200,229 sentences, 166.46M tokens)
  ├── 1. lowercase Cyrillic regex tokenization (preserving 9 Bashkir letters)
  ├── 2. enforce strict sentence boundaries (n-grams never cross sentence boundaries)
  ├── 3. compile unigrams from canonical frequency index (921,860 wordforms)
  ├── 4. slide 2-word windows and aggregate bigram frequencies
  ├── 5. slide 3-word windows and aggregate trigram frequencies
  ├── 6. prune singleton sequences (minimum cutoff: count >= 2)
  └── 7. export zstd Parquet: unigrams (921,860), bigrams (13,097,446) & trigrams (19,888,009)

Quality and Use

This is a statistical index, not a normative Bashkir dictionary or a grammar checker. At this scale, words and phrases from other languages, borrowings, names, regional vocabulary, technical terminology, OCR artifacts and other noise may remain; this is normal for a large web-derived corpus. Always validate frequency based suggestions before using them in production, spellchecking, linguistic research or a user-facing application.

Limitations

• Presence in an n-gram file is not proof that a word or phrase is standard Bashkir.
• Frequent sequences can reflect names, boilerplate or segmentation artifacts.
• Counts reflect the underlying corpus composition, not usage norms.
• For OCR and corpus cleaning, combine n-gram scores with language identification, dictionaries and human review.

Related Resources

• Bashkir Frequency Index — standalone ranked word frequencies; intended to be used together with this release, which adds within-sentence word sequences.

Usage

pip install datasets pandas

from datasets import load_dataset

bigrams = load_dataset("failed09/bashkir-ngram-index", "bigrams")
print(bigrams["train"].head())

Or directly with Pandas:

import pandas as pd

unigrams = pd.read_parquet("unigrams.parquet")
bigrams = pd.read_parquet("bigrams.parquet")
trigrams = pd.read_parquet("trigrams.parquet")

For large-scale processing, read only the columns you need and use Parquet filters or streaming batches where supported.

License

Distributed under the CC BY 4.0 license. The release contains derived statistics, not the source texts. Upstream source licenses and attribution requirements still apply to the underlying materials.

Citation

@dataset{failed09_bashkir_ngram_index_2026,
  title = {Bashkir Word N-gram Index v11.6},
  author = {failed09},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/failed09/bashkir-ngram-index},
  note = {Open-source Bashkir word n-gram index for corpus processing and linguistic research}
}

Open Bashkir Data and Sources 🐝

This release is part of an open-source effort to support the development, preservation and practical use of the Bashkir language. Other related models, datasets and tools are available on the author's Hugging Face profile.

The author does not claim ownership or authorship of the source texts or other materials used to derive this release; rights and licensing remain with the original authors, publishers and dataset providers. Source texts are not redistributed in this repository, so users should follow the licenses and attribution requirements of the relevant upstream resources.