failed09/bashkir-periodicals-izddom

Bashkir Periodicals Cleaned (Izddom)

High-quality cleaned and structured Bashkir periodicals (Respublika Bashkortostan Publishing House) for LLM pretraining and fine-tuning.

Overview

A curated, deduplicated and filtered edition of modern Bashkir periodicals derived from the original bashkorttele/periodicals-izddom dataset published by the Foundation for the Preservation and Development of the Bashkir Language.

The dataset includes material from 13 leading Bashkir-language journals and newspapers (including Amanat, Shonkar, Vatandash, Akbuzat, Yeshlek, Bashkortostan Kyzy, and others). Non-text puzzles, crosswords, editorial boilerplates, and page headers were systematically removed, and text quality was verified using neural language identification.

At a glance
Task	Text generation, Fill-mask, Language Modeling
Default config	articles (37,870 articles, 22.08M words)
Fields	id, collection, text, n_words, n_chars, n_sentences, ba_prob, source_id
Source	Respublika Bashkortostan Publishing House (via bashkorttele/periodicals-izddom)
License	CC-BY-4.0

Contents

Files and Configurations

Config	Contents	Rows / Size
articles (recommended)	Full cleaned articles for LLM pretraining	38,039 rows (86.2 MB)
sentences	Deduplicated sentences (3–40 words) for fine-tuning	1,665,735 rows (73.5 MB)
META.json	Release passport: version, upstream source, and artifact hashes	—
SHA256SUMS	SHA-256 checksums for public files	—

Schema

articles configuration (default)

Column	Type	Description
id	string	Normalized identifier (izddom-clean-XXXXX)
collection	string	Publication name (e.g. Аҡбуҙат, Йәшлек, Шоңҡар)
text	string	Clean coherent article text with normalized hyphens and stripped boilerplate
n_words	int64	Word count of the cleaned article
n_chars	int64	Character count of the cleaned article
n_sentences	int64	Number of extracted sentences in the article
ba_prob	float64	Neural Bashkir language probability from failed09/bashkir-lid
source_id	string	Original upstream record identifier

sentences configuration

Column	Type	Description
sentence_id	string	Unique sentence identifier ({source_id}-sXXX)
text	string	Clean individual sentence
collection	string	Publication of origin
word_count	int64	Sentence word count
source_id	string	Upstream article identifier

Examples

Articles configuration:

Collection	Preview	ba_prob	Words
Акбузат	Һөйөклө “Аҡбуҙат”! 1991 йылда исем туйыңды үткәргәндә Нурмөхәммәт мөфти хәҙрәт һиңә аҡ юл теләп фатихаһын биргәйне...	1.0000	183
Шонкар	Башҡорт халҡының тарихында рухи мәҙәниәттең тотҡан урыны ғәйәт ҙур. Халҡыбыҙ быуаттар буйына үҙенең милли йөҙөн һаҡлап ҡалған...	0.9998	412

Sentences configuration:

Collection	Text	Words
Акбузат	Кейеп ебәр кәзәкей.	3
Акбузат	Паркка киттек, күрегеҙ.	3
Шонкар	Донъяға яңы китап пәйҙә булды.	5

Method

upstream jsonl (39,711 records)
  ├── 1. strip headers & page numbers (e.g. "26 Башҡортостан ҡыҙы №5 2022")
  ├── 2. reject crosswords, scanwords & horoscopes (1,261 dropped)
  ├── 3. strip trailing editorial markers ("Дауамы бар", illustrator credits, author lines)
  ├── 4. normalize whitespace and detached OCR punctuation ("һүҙ ." -> "һүҙ.")
  ├── 5. anonymize PII: mask bank cards ([КАРТА]), phones ([ТЕЛЕФОН]), emails ([EMAIL])
  ├── 6. neural language filter: failed09/bashkir-lid (ba_prob >= 0.90) (411 dropped)
  ├── 7. sentence segmentation with initials/abbreviation protection, casing & syntax filters, deduplication (1,665,735 unique sentences)
  └── 8. export zstd Parquet: articles (38,039) & sentences (1,665,735)

Quality and Use

Data Sanitization & PII Protection

To protect personal privacy, all texts were passed through automated PII detection. Phone numbers, bank card numbers (frequently found in public charity appeals), and email addresses were scrubbed and replaced with unified placeholder tokens: [КАРТА], [ТЕЛЕФОН], and [EMAIL].

Limitations

• Fully Automated Pipeline: Text extraction, boilerplate stripping, language identification, and sentence splitting were executed entirely automatically. Consequently, residual OCR artifacts from upstream PDF digitization, broken hyphenations, rare fragmented phrases, or imperfect sentence boundary detections may remain.
• Dialect and Register Variation: Material spans diverse periodicals (from children's magazines like Акбузат to cultural and literary publications like Шонкар and Агидель), including regional dialects, literary idioms, and colloquial speech.
• Formatting and Layout: Magazine formatting (poetry verses, multi-speaker dialogues, numbered lists) is preserved with original line breaks where appropriate.

Related Resources

• Bashkir Books & Folk Literature (Kitap) — companion dataset containing 5,641 cleaned sections and 73.8k sentences from literary editions, folklore, and cultural publications by Kitap Publishing House.
• Bashkir LID (Binary BA vs NON_BA) — compact ONNX language gate used for neural language filtering (ba_prob >= 0.90) in this dataset.
• Bashkir-Russian Phrasebook Corpus — reviewed and cleaned conversational parallel corpus, also derived and optimized from bashkorttele releases.

Usage

pip install datasets pyarrow

from datasets import load_dataset

# Load full articles for pretraining
dataset = load_dataset("failed09/bashkir-periodicals-izddom", "articles", split="train")
print(dataset[0]["text"][:200])

# Load deduplicated sentences
sentences = load_dataset("failed09/bashkir-periodicals-izddom", "sentences", split="train")
print(sentences[0])

License

This dataset is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, matching the license of the upstream materials.

• Upstream Dataset: bashkorttele/periodicals-izddom
• Rights Holder of Source Publications: State Unitary Enterprise Publishing House "Respublika Bashkortostan"
• Publisher of Upstream Compilation: Autonomous Non-Profit Organization "Foundation for the Preservation and Development of the Bashkir Language" (bsfond.ru)
• Cleaning, LID Verification and Segmentation: failed09 / BashkirCorpus

Citation

@dataset{failed09_bashkir_periodicals_izddom_2026,
  title     = {Bashkir Periodicals Cleaned (Izddom)},
  author    = {failed09},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/failed09/bashkir-periodicals-izddom},
  note      = {Curated, cleaned and LID-verified Bashkir periodicals based on bashkorttele/periodicals-izddom (CC BY 4.0)}
}

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.