Til-Books-KazNEB

Kazakh (and some Russian) book texts harvested from KazNEB — the Kazakhstan National Electronic Library (kazneb.kz) — and turned into machine-readable text via PyMuPDF text-layer extraction or Qwen3-VL OCR of page scans. Part of the TilQazyna Kazakh LLM corpus program.

TL;DR

• Two configs. processed = quality-filtered, judged real books (use this for training). raw = the full harvest including previews/garbage (audit/debug only).
• Each book is one row: full text + metadata + a model-assigned quality score and tier.
• Previews are filtered out. KazNEB's median item is only 4 pages — most IDs expose just a cover, table-of-contents, or a restricted preview. We keep only items with ≥10 pages and drop degenerate OCR.

How it was built

1. ID discovery. KazNEB's catalogue caps deep pagination at ~10k results, so we enumerate the whole item-ID space directly (catalogue/view/<id>): a real item has a real <title>; placeholder IDs render Book #<id>. The last real ID is ~1.76M.
2. Harvest (per ID). catalogue/view → a signed full.pdf text-layer (extracted with PyMuPDF) when present; otherwise the page images from bookView are downloaded and OCR'd with Qwen3-VL (Kazakh-aware prompt, verbatim transcription).
3. Quality gate (at harvest). Items with < 10 pages (cover / TOC / restricted preview) or degenerate OCR (one or two characters dominate, e.g. unreadable old Arabic/Latin scans) are dropped before OCR — saving compute and keeping the corpus clean.
4. Judge. Every kept book is scored by a Qwen judge → {score 1–5, category, lang}, plus a character-level kk_ratio. tier = premium (score≥4 & kk_ratio≥0.5) / clean (score≥3) / raw.

Why filtering matters (measured)

A 1000-book sample, by page count:

pages	share	what it is
1–2	36 %	cover only / not-a-book
3–9	29 %	preview / table-of-contents
10–49	12 %	short books
50–199	13 %	real books
200+	9 %	full books

~65 % of KazNEB IDs are previews/covers (median 4 pages). The page-count + garbage filter removes this junk so the processed config contains real, readable books only.

Schema

processed config (use for training):

field	type	meaning
id	string	KazNEB item ID
title	string	catalogue title (author / library suffix included)
year	string	year parsed from title, if any
text	string	full book text (digital extraction or OCR)
text_raw	string	pre-judge text (when a clean layer was applied)
n_pages	int	pages extracted/OCR'd
route	string	textpdf (digital text layer) or ocr (scanned)
score	int	judge quality 1–5
category	string	judge topic label
judge_lang / kk_ratio	string / float	language label / Kazakh-character ratio
tier	string	premium / clean / raw
source	string	kazneb

Usage

from datasets import load_dataset
ds = load_dataset("TilQazyna/Til-Books-KazNEB", "processed", split="train")
premium = ds.filter(lambda r: r["tier"] == "premium")          # cleanest Kazakh book text
kk = ds.filter(lambda r: r["judge_lang"] == "kk")

Known limitations

• OCR noise. Scanned books carry OCR artifacts (page numbers, running headers, hyphenation breaks сло-\nво, occasional misreads). The judge tolerates these for genuine prose, but the text is not yet model-cleaned — a dedicated clean pass is planned. The score reflects readability.
• Multilingual. KazNEB holds Kazakh, Russian and some foreign-language items. Use judge_lang / kk_ratio to select Kazakh-only.
• Truncated items. A few books expose only some pages on KazNEB itself; we keep what is available.

Provenance & ethics

Sourced from the public KazNEB catalogue for non-commercial Kazakh-language research. Respect the rights of the source library and original authors. Restricted/preview-only items are intentionally excluded.