Rudaw — Sorani news corpus (rudaw)

23,758 Central Kurdish (Sorani) articles from rudaw.net, each with its category and canonical URL.

At a glance

Rows	23,758 (single train split)
Columns	text (string), category (string), url (string)
Parquet on disk	39.7 MB (96.1 MB uncompressed)
Row groups	24
Script	Arabic script; Latin characters in ~18% of rows (URLs, names, agencies)
Language	Central Kurdish / Sorani (ckb)

Field profile

Measured over a 1,000-row sample (first row group):

Column	Mean	Median	Max	Notes
text	4,530 chars	2,086	110,415	often begins with a leading newline
category	7.4 chars	7	9	Sorani label, e.g. هەڤپەیڤین (interview)
url	47.1 chars	47	57	https://www.rudaw.net/sorani/<section>/<id>

Sample row:

text     : "…گوتەبێژی بەرپرسی سیاسەتەکانی دەرەوەی یەکێتی ئەوروپا رایگەیاند، ئەوان ناڕازین…"
category : "هەڤپەیڤین"
url      : "https://www.rudaw.net/sorani/interview/130620231"

Usage

from datasets import load_dataset

ds = load_dataset("razhan/rudaw", split="train")
docs = ds.filter(lambda r: r["category"] == "هەڤپەیڤین")   # interviews only

category and url make this corpus convenient for domain-balanced sampling and for re-crawling/expanding the same sections.

How it was built

Site crawl of the Sorani section of rudaw.net; article text extracted per page, with section name kept as category and the source URL as url.

Known limitations

• Leading/trailing whitespace and newlines are present in text; strip before use.
• Boilerplate (related-article teasers, video captions, share text) survives in some rows.
• No deduplication and no date field — ordering information lives only in the URL.
• Single-publisher corpus: same editorial bias caveat as any newspaper crawl.

Related

• razhan/anfsorani — a second Sorani news crawl (article text only).