TilQazyna / Til-Books

A large corpus of full-text books in Kazakh (and Russian/English co-resident in the same collections), assembled by extracting and OCR-ing digitised book holdings. 16,847 books, each provided in two layers: the exact extracted/OCR text (text_raw) and a cleaned, reflowed reading text (text).

Part of the TilQazyna master collection.

---

Abstract

Til-Books turns scattered digitised book holdings — born-digital PDFs, scanned page images, and mixed-format archives — into a single, schema-uniform, quality-tiered book corpus for Kazakh language modelling. Born-digital PDFs are text-extracted directly; image-only / scanned pages are recognised with a vision LLM (Qwen3-VL) under a Kazakh-tuned OCR prompt. Every book is stored twice: text_raw preserves the exact extraction/OCR (book layout intact), and text is the same content with book artifacts removed (de-hyphenation, paragraph reflow, page-number/​header/​footer stripping) — a rule-based transform that never rewrites the words, so there is no model hallucination. Books are tiered premium / clean / raw by language, length, and cleanliness.

At a glance

Books	16,847 (deduplicated from 17,510 ingested; 663 near-duplicates dropped)
Layers	text_raw (exact extract/OCR) + text (cleaned, reflowed)
Tiers	premium 12,988 · clean 2,593 · raw 1,266
Languages	kk 13,194 · ru 2,076 · en 1,209 · mixed 355 · other 13
Sources	digitised-PDF 13,848 · scanned-PDF (OCR) 1,575 · 600-book set 1,050 · KazNEB scans (OCR) 374
OCR engine	Qwen3-VL-30B-A3B (vision), Kazakh OCR prompt
License	CC-BY-4.0 · Access: gated (manual)

Why this dataset

There is no clean, full-text Kazakh book corpus on the open hub — the large Kazakh corpora are web/news, and digital libraries (KazNEB, etc.) serve only scanned page images behind viewers. Books are the highest-quality long-form register (edited prose, literature, science, law), which is exactly what is scarce for Kazakh LM training. Til-Books fills that gap with both the raw OCR (for those who want fidelity) and a cleaned reading text (for training).

Sources & provenance (source column)

source	what	text path
staged-pdf	born-digital PDFs from a national document archive	PyMuPDF text-layer extraction
staged-scanned	image-only PDFs from the same archive	Qwen3-VL OCR (per page)
600kitap	the «Қазақша 600 кітап» collection (mixed formats)	direct extraction + OCR
kazneb	scanned books from KazNEB (page-image folders)	Qwen3-VL OCR (per page)

Companion repos with additional book sources harvested separately: Til-Books-Adebiportal, Til-Books-InternetArchive.

Two-layer text — what text vs text_raw mean

• text_raw — the exact extracted / OCR'd text. Faithful to the page (line breaks, hyphenation, page numbers, running headers all present). Use this if you need maximum fidelity or want to run your own cleaning.
• text — the same content with book layout removed by rules (no LLM, no rewriting):
  • de-hyphenation across line breaks (сло-\nво → слово)
  • paragraph reflow (hard single line-breaks joined; blank line = paragraph boundary)
  • standalone page numbers, recurring running headers/footers, and УДК/ББК/ISBN boilerplate stripped
  • whitespace normalised Use this as ready-to-train reading text.

Schema

column	type	description
text	string	cleaned, reflowed reading text (layer 2)
text_raw	string	exact extracted / OCR text with book layout (layer 1)
source	string	staged-pdf / staged-scanned / 600kitap / kazneb
book_id	string	source identifier of the book
n_chars	int	length of text
n_chars_raw	int	length of text_raw
lang	string	detected dominant language (kk/ru/en/mixed/other)
tier	string	premium / clean / raw

Quality tiers (cumulative)

Tiered by cheap signals (language, length, alphabetic-character ratio):

• premium — Kazakh, ≥ 1000 chars, clean ratio ≥ 0.7 (well-extracted Kazakh books)
• clean — premium plus kk/ru/mixed, ≥ 300 chars, clean ratio ≥ 0.55
• all — everything, including short / noisy / heavily-OCR-degraded books

clean reads premium + clean shards; all reads everything. Single-copy storage via config globs (no duplication on disk).

Usage

from datasets import load_dataset

# cleanest Kazakh books, ready-to-train cleaned text
ds = load_dataset("TilQazyna/Til-Books", "premium", split="train")
print(ds[0]["text"][:1000])          # layer 2 (cleaned)
print(ds[0]["text_raw"][:1000])      # layer 1 (exact OCR/extract)

# broader set
ds = load_dataset("TilQazyna/Til-Books", "clean", split="train")

# everything
ds = load_dataset("TilQazyna/Til-Books", "all", split="train")

kk_books = ds.filter(lambda r: r["lang"] == "kk")

Limitations & biases

• OCR error — kazneb and staged-scanned books are machine-recognised (Qwen3-VL) and contain recognition errors; prefer premium for cleaner text, or use text_raw to assess.
• Cleaning is heuristic — the text reflow is rule-based; on unusual layouts (poetry, tables, multi-column) it can over- or under-join lines. text_raw is always the safe fallback.
• Mixed languages — book collections contain Russian and some English; lang tags the dominant language per book, not per passage.
• Tiering is signal-based, not human-judged.

Ethical & legal

• Assembled from digitised library/archive holdings for Kazakh-language NLP research.
• Copyright status varies by book and year; gated (manual) to match org policy and to keep access deliberate. Do not redistribute in-copyright full texts outside research use.
• No targeted de-anonymisation; incidental personal data in books must not be used to profile individuals.

Citation

@misc{tilqazyna_books,
  title  = {TilQazyna Til-Books: a two-layer Kazakh full-text book corpus},
  author = {TilQazyna},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/TilQazyna/Til-Books}}
}

FAQ

Why two text columns? Fidelity vs usability. text_raw is exactly what came off the page; text is the cleaned reading version. You choose; nothing is invented.

Was the text rewritten by an LLM? No. Cleaning is pure rules (de-hyphenate, reflow, strip page furniture). An LLM (Qwen3-VL) was used only to OCR scanned images into text_raw — not to paraphrase.

How many are real Kazakh books? 13,194 are tagged kk; the rest are Russian/English/mixed co-resident in the same archives. Filter on lang.

Which tier for pre-training? premium for cleanest Kazakh; clean for a broader mix; all if you filter yourself.

Is this everything? No — it is the digitised holdings processed so far (national archive PDFs + KazNEB + the 600-book set). KazNEB alone has ~40k books; harvesting continues. See the companion Til-Books-* repos for additional sources.

Why gated? Org policy + copyright caution. Access is granted on request.