Datasets:
Request access to the boolak corpus
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This corpus aggregates sources with heterogeneous provenance: public crawls, Tatar-language press, and digitised books. It is made available for research on low-resource languages. By requesting access you agree not to redistribute it and not to use it commercially.
Log in or Sign Up to review the conditions and access this dataset content.
Ultimate cleaned Tatar Corpus
The training corpus of the boolak project (a Tatar language model trained from
scratch), frozen right before tokenization. One row = one document. The text
is cleaned, filtered and deduplicated, but not tokenized: this is the exact
input of to_ids.py.
- Version:
v20260812— built on 2026-08-07 - 954,382 documents · 2.52 billion characters
- 526 million tokens (project tokenizer
tokenizer_32K.json, excluding<eos>) - 15 sources, drawn from 133 files
from datasets import load_dataset
ds = load_dataset("eldiablo92/bulak-ultimate-tatar-corpus", split="train", revision="v20260812") # pin the version
ds = load_dataset("eldiablo92/bulak-ultimate-tatar-corpus", split="train", streaming=True) # without downloading
ds.filter(lambda r: r["source"] == "books_w3") # a single source
Access
The corpus is public but gated: this card, the statistics and the composition below are open to everyone, while the files require approval. Request access from the button at the top of this page.
Columns
| column | type | description |
|---|---|---|
id |
string | <origin>:<index>, unique across the corpus |
source |
string | provenance group (see table below) |
origin |
string | exact source file, before grouping |
text |
string | document, NFC-normalized |
n_chars |
int32 | length in characters |
n_tokens |
int32 | length in tokens |
Tatar is written with composed letters (ә, ө, ү…) that arrive either precomposed or decomposed depending on the source. Everything is in NFC: without it, two texts that look identical hash differently and deduplication misses them.
Sources
source groups the individually digitised books together; origin keeps the
exact file for every document, so nothing is lost.
| source | files | documents | characters | tokens | chars/token |
Pipeline
- Extraction — PDF/EPUB/FB2 from waves 1–3 (encyclopedia, monographs, literature) to text; scraped press and public crawl datasets.
- Cleaning — boilerplate, headers/footers, non-Tatar fragments, OCR artefacts.
- Quality filtering — Gopher heuristics, quality classifier, GlotLIDlanguage identification, HPLT-specific filter.
- Deduplication — exact, then MinHash across documents, then repeated lines across sources.
- OCR lexical repair , driven by a lexicon built from the deduplicated corpus.
Limitations
- Proportions are heavily imbalanced: crawl dominates by volume, while literary and encyclopedic sources dominate by quality. Reweight as needed.
- Despite filtering, residual Russian remains in the crawl-derived sources.
- Digitised books go through OCR: residual errors are rare but non-zero, especially on older orthographies.
- Downloads last month
- 35