Kteb — 1,000+ OCR'd Kurdish books

1,033 Kurdish books as full text, produced by OCR'ing scanned books. Each row is one book: its Markdown-formatted text plus the book's title.

At a glance

Rows	1,033 books
Columns	text (full book, Markdown), book_title (string)
Parquet on disk	77.5 MB (157.0 MB uncompressed)
Row groups	2
Language	Central Kurdish / Sorani (mostly), Arabic and some Hawrami
Measured total	12,240,756 words across the 1,033 books

Field profile

Measured over the whole dataset:

Mean book length	82,647 characters (~11,850 words)
Median book length	71,622 characters
Longest book	308,137 characters
Mean title length	30 characters
Books whose text uses Markdown headings (#, ##)	923 of 1,033 (89%)

Sample titles:

(سەعدى) شاعير فارسێکى شيرازيى، يان کوردێکى هەورامى؟
10 بنەما بۆ پاککردنەوەی دەروون
(جەزیری) ئەندازیاری کورد باوکی ئەندازیاری ڕۆبۆتێک لە جیهان
100 پرسیار و وەڵام لەسەر هەڵەکانی عومرەکاران

Sample text (truncated):

# (۱۰) بنەما بۆ پاککردنەوەی دەروون

نووسینی:
عبدالرزاق بن عبدالمحسن البدر وەرگێڕانی:
دەرباز گەرمیانی ١٤٤١٨ - ٢٠٢١ز

Note the collector's earlier stated figures (12,484,235 words, 1,565,055 unique word forms) versus the 12,240,756 words measured on this snapshot — the difference comes from how titles and repeated whitespace are counted, so treat word counts as approximate.

Usage

from datasets import load_dataset

ds = load_dataset("razhan/kteb", split="train")
print(ds[0]["book_title"])
print(ds[0]["text"][:1000])

Rows are whole books, so stream or filter by title rather than pulling everything into memory at once.

How it was built

The books were scanned and OCR'd into Markdown (<title>.md per book, one directory per title). The CSV that became this dataset was produced by create_csv.py from the private razhan/books archive, which:

1. removes LaTeX fragments ($…$, \[…\], \command{…}) and repeated-character words, and
2. applies a resolve_ae() normalisation that standardises the Kurdish ه/ە pair word-initially and word-finally.

Known limitations

• OCR noise. Character-level errors, dropped diacritics, and broken lines follow the quality of the scanned pages; older or low-quality scans are worse.
• Mixed orthography. Titles and text mix Arabic/Persian forms (ي, ك, ه) with standard Sorani (ی, ک, ە), and some text carries stray Hebrew letters (leftovers from transliteration fonts). Normalise with razhan/script_normalization_ckb before training a tokenizer.
• Markdown structure is inconsistent: 110 books have no heading markup at all.
• Page numbers and headers/footers from the scans are not stripped.
• Rights: these are scanned in-copyright books. The corpus is fine for research use; check each title before redistributing text.
• No metadata beyond the title — no author, year, or page count (use razhan/kteb-dataset, which adds source, num_pages and num_words).

Related

• razhan/books (private) — the raw 7z archive and the scripts that produced this CSV.
• razhan/kteb-dataset — 10,124 books with page and word counts.
• razhan/ktr — word images rendered from Kurdish book text for OCR training.