You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Kurdish News Corpus

3.42M news articles crawled daily from 259 Kurdish and Kurdish-regional news websites between 2026-02-10 and 2026-11-13, built for Kurdish NLP work (language modeling, dialect classification, low-resource text mining). Kurdish is split across two main standard varieties β€” Sorani (ckb, Arabic script) and Kurmanji (kmr, Latin script) β€” plus several closely related regional varieties this crawl organically picked up: Southern Kurdish / Pehlewani (sdh), Hawrami (hac), and Zazaki (zza).

This is the final, deduplicated snapshot of a daily-crawling project that has been shut down; it will not be updated further.

Dataset summary

Articles 3,422,229
Total words ~1.01 billion
Distinct source domains 259
Date range 2026-02-10 – 2026-11-13
Format single Parquet file, ZSTD-compressed
Columns 47 (see schema below)

Dialect / variety breakdown (dialect column)

dialect rows share
sorani 1,783,994 52.1%
kurmanji 781,697 22.8%
(empty) 330,238 9.6%
unknown 252,005 7.4%
southern (sdh) 185,888 5.4%
hawrami (hac) 59,191 1.7%
mixed 11,014 0.3%
bahdini 10,247 0.3%
zazaki (zza) 7,711 0.2%
sorani_arab / kurmanji_latn 244 <0.01%

dialect is computed from arabic_script_ratio (>0.8 β†’ sorani, <0.2 β†’ kurmanji, else mixed) independently of the language column below, so it's populated even on older rows that predate language-ID tagging.

Top sources by article count

rudaw.net (229,348), wishe.net (188,982), westganews.net (176,314), xendan.org (173,945), sharpress.net (163,666), speemedia.com (122,602), cinarinsesi.com (113,854), kurdistantv.net (100,501), lalishduhok.com (97,681), nrttv.com (92,310), 964.krd (60,389), knnc.net (58,216), yariga.net (57,997), sorani.anf-news.com (57,446), shafaq.com (51,986), and 244 more.

Schema

47 columns, the same layout the source project used for its live Parquet corpus. Key fields:

  • url, canonical_url, source_domain, source_name β€” provenance
  • title, body, description, author, authors, category, tags, image_url β€” article content/metadata
  • publish_date, modified_date, crawl_date β€” dates (string, mixed YYYY-MM-DD / full ISO-timestamp formats across the crawl's lifetime)
  • word_count, char_count, sentence_count, line_count, mean_word_length, terminal_punct_ratio, short_line_ratio, duplicate_line_frac, symbol_to_word_ratio, curly_bracket, arabic_script_ratio, quality_score β€” text-quality signals (the kind of features used by corpora like C4/OSCAR to filter boilerplate)
  • language, lang_label, lang_confidence, is_kurdish, source_language, dialect, script β€” language/dialect ID
  • content_hash β€” dedup key used by the source project
  • section, keywords_site, iptc_topic, iptc_topic_score, genre, entities_persons, entities_organizations, entities_locations, sentiment, sentiment_score β€” enrichment fields (sparsely populated; not every row went through every enrichment stage)
  • extraction_method, crawl_type β€” pipeline provenance (id is an internal row identifier)

How this file was built

Merged from 285 daily-partitioned Parquet files (the source project wrote one file per crawl day, plus occasional re-extraction passes after a parser fix) into one file, deduplicated by URL β€” when an article was re-extracted after a site's HTML selectors were fixed, only the latest version is kept here (689,695 superseded rows dropped). One domain, evronews.net (3 rows), was excluded after manual review found it to be a domain-squatter hijack serving unrelated Chinese-language gambling spam under a Kurdish-sounding name β€” no other exclusions were made.

Known limitations β€” please read before using

  • ~15% of rows are not core Kurdish text. The language column (from an NLLB fastText language-identification model) shows: ckb 1,834,499, kmr 765,638, sdh 185,888, (empty) 206,194, ar (Arabic) 179,141, eng (English) 79,549, hac 59,191, plus smaller counts of mzn, ug, pa and others. The Arabic/English rows are largely legitimate β€” Kurdish outlets often publish bilingual wire content, and Sorani (Arabic script) is occasionally misclassified as Arabic by the LID model β€” but if you need a strictly-Kurdish subset, filter on language in {ckb, kmr, sdh, hac, zza} rather than assuming every row is Kurdish.
  • 2.6% of rows (88,482) come from domains outside the project's curated 344-source list, picked up by opportunistic domain discovery. By volume these are mostly legitimate (ANF's English edition, a Hadith reference site's Kurdish-translation section, several Kurdish outlets with domain-string drift from the curated config) rather than spam.
  • 6% of rows (206,010), all from curated/known-good domains, have an empty language field β€” they predate the project adding LID tagging. Their dialect field (computed independently) is still populated.
  • Text-quality fields (quality_score, duplicate_line_frac, etc.) were computed by the source pipeline's own heuristics and were not independently re-verified for this export β€” treat them as a useful starting filter, not ground truth.

License / usage

The underlying article text is copyrighted by its original publishers β€” this project collected it for Kurdish NLP research and makes no copyright claim over the content itself. No explicit redistribution license is asserted here (marked license: other deliberately, not a permissive license this project isn't entitled to grant). If you plan to use this for anything beyond research β€” training, redistribution, commercial use β€” verify your own rights/fair-use basis per source outlet. This follows the same "unspecified, research-use" convention used by comparable scraped-news corpora (e.g. CC-News).

Citation

No formal citation target; if you use this, a link back to this dataset page is appreciated.

Downloads last month
76