muzaffercky/kurdish-ins

Kurmanji Kurdish Instructions

32,991 instruction–response pairs in Kurmanji Kurdish (kmr_Latn). Each pair has a question and an answer about a passage of real Kurmanji text. Gemini wrote the questions. The answers are not generated: each one is copied word for word from its passage, so every answer is text that Kurmanji speakers wrote.

Subsets

Subset	Pairs	What
kmr_Latn_clean	30,927	The pairs with garbled letters repaired or dropped. Use this one for training
kmr_latn	32,991	The pairs as they were generated, with the garbled letters of their sources

kmr_Latn_clean was used to instruction-tune Sorjin2.1-12B-Instruct.

The pipeline that built it is at github.com/cikay/kurdish_ins.

Loading

from datasets import load_dataset

ds = load_dataset("muzaffercky/kurdish-ins", "kmr_Latn_clean", split="train")

Both subsets have only a train split.

Example

{
  "instruction": "Li ser jiyan û kariyera helbestvan û nivîskar Sezaî Karakoç çi agahî tê zanîn?",
  "output": "Sezaî Karakoç (jdb. 22ê kanûna paşîn a 1933, Erxanî - 15 Mijdar 2021), helbestvan, nivîskar û siyasetmedarê tirk e.\n\nZarokahîya xwe li Erxanîyê, madenê û li Dîcleyê derbas bûye. Li Mereş û Intabê perwedehîya xwe domand.\n\nHelbesta wî ya nav Mona Rosa gelek navdar e. bi eslê xwe kurd e",
  "context": "…",
  "id": "<urn:uuid:…>",
  "domain": "ku.wikipedia.org",
  "source": "HuggingFaceFW/fineweb-2",
  "url": "https://ku.wikipedia.org/wiki/Seza%C3%AE_Karako%C3%A7"
}

Fields

Field	Description
instruction	A question in Kurmanji, written by Gemini to stand on its own without the passage
output	The answer: one or more spans copied verbatim from context, joined in their order
context	The passage the question was asked about and the answer was copied from
id	The id of the source document in its dataset: the FineWeb-2 id, or the thesis number
domain	The website the text came from (tez.yok.gov.tr for the theses)
source	The Hugging Face dataset the text came from
url	The page the text came from

Lengths in words:

	Min	Median	Mean	Max
instruction	5	22	23	60
output	16	196	223	1,392
context	50	784	827	5,682

Sources

Source	Pairs	Documents	License
HuggingFaceFW/fineweb-2 (kmr_Latn)	21,270	19,032 web pages from 1,117 domains	ODC-By 1.0
muzaffercky/kurdish-theses-yok	11,721	408 theses	CC BY 4.0

The largest web domains are ku.wikipedia.org (5,989 pairs), ku.m.wikipedia.org (1,223), krd.sputniknews.com (888), krd.riataza.com (816), www.bernamegeh.org (673) and candname.com (559).

How it was made

1. Classify. Every FineWeb-2 Kurmanji page and every thesis was labelled with its web register (encyclopedia, news, opinion, how-to, machine translation…) by TurkuNLP/web-register-classification-multilingual.
2. Select. Web pages were kept when they scored above 0.5 on an informational register (informational, encyclopedia, biography, FAQ, how-to, recipe) and below the threshold on every register whose facts go out of date or are not facts: news, sports, blogs, opinion, promotion, ads, and machine translation (0.2, stricter, to drop spam sites auto-translated into Kurmanji). Kurdish Wikipedia, Wiktionary and Kurdipedia skip only the machine-translation check, since many of their articles score 0.2–0.3 on it. Every thesis was kept.
3. Chunk. Texts longer than 1,500 words were cut at line breaks into chunks of at most 1,500 words; a text with no line breaks stays whole, which is why some contexts are longer. Chunks under 50 words were skipped.
4. Generate. gemini-3.8-flash (temperature 0.3, thinking level low, Batch API) read each chunk and wrote one self-contained, open-ended question in Kurmanji, plus the spans of the chunk that answer it. The prompt asks for whole sentences, as long an answer as the question allows, and no reference to "the text" or "the author", since the question is shown without the passage in training. The prompt is in kurdish_ins/prompts.py.
5. Verify. A pair was kept only when every span appears verbatim in the chunk, in order and without overlapping, and the joined answer has at least 15 words. Spans are joined by a blank line where a line break was skipped and by a space elsewhere.

Of 37,601 chunks, 33,073 became pairs:

Outcome	Chunks
Kept	33,073
A span was not found verbatim in the chunk	2,969
Gemini found no question with a long, complete answer	1,531
Reply was not valid JSON	26
Answer under 15 words	2

6. Remove contact details. 82 pairs whose answer or context has a phone number or an email address were dropped. FineWeb-2 replaces emails with placeholders (firstname.lastname@example.org, email@example.com, kept here) but not phone numbers, and the theses end with their authors' contact pages.

Cleaning (kmr_Latn_clean)

The thesis PDFs and some FineWeb-2 pages have garbled letters, and a model trained on them learns to write them. kmr_Latn_clean is kmr_latn after scripts/clean.py, with the settings in clean.yml. Text is put in Unicode NFC first.

Replaced, where the correct letter is certain (25,606 letters). The question, the answer and the context get the same replacements, so every answer is still a verbatim copy of its context:

Garbled	Correct	Cause	Letters
þ Þ	ş Ş	Turkish text read in the wrong code page (cp1254 as cp1252): piþtî	19,811
ș Ș	ş Ş	Romanian look-alike, another character to a tokenizer	2,804
ȇ ȋ ȗ	ê î û	Written with an inverted breve: zimanȇ	1,994
ÅŸ ÄŸ	ş ğ	UTF-8 decoded twice	515
ҫ Ҫ	ç Ç	Cyrillic look-alike	482

Dropped, a pair whose question or answer has garbling with no single correct form (2,064 pairs). The context is not checked: it is not the text a model learns to write, and checking it would drop many more pairs.

Pattern	Example	Pairs
Ģ ģ ġ Ġ from the thesis PDFs	PiĢtî; ġ is sometimes a broken ş, sometimes a transliteration letter	1,325
ì for î in the thesis PDFs	jì, Kurdì	293
Footnote numbers turned into subscripts	₁	117
― and ‖ for quotation marks in the thesis PDFs	―Ez … ‖	79
Kurmanji typed with Turkish ı	jı, bı, lı	77
Symbols from the PDFs' own fonts (Unicode private use area)	invisible as text	65
ý Ý ð: Turkish ı İ ğ in the wrong code page, but also real letters of Czech and Icelandic names	Heyrovský	51
Symbols some theses use for transliterated Arabic letters, and @	Mu√ammed, Fa¡ilatun	35
è for the Arabic ayn before a vowel or capital	Seèîd	14
ÿ	ÿzdîya	8

Kept by source:

Source	kmr_latn	kmr_Latn_clean
HuggingFaceFW/fineweb-2	21,270	21,113
muzaffercky/kurdish-theses-yok	11,721	9,814

Most dropped pairs come from the theses, so the clean subset leans a little less towards academic text.

Language check

GlotLID was run on every question and answer before the contact details were removed (33,073 pairs):

	Pairs
Question not identified as kmr_Latn	770 (kiu_Latn 764, diq_Latn 6)
Answer not identified as kmr_Latn	110 (kiu_Latn 64, diq_Latn 44, others 2)

No pair was dropped for this. The questions GlotLID called kiu_Latn (Kirmanjki) that were read are plain Kurmanji; GlotLID seems to confuse short Kurmanji text with Zazaki varieties. Some of the diq_Latn answers are genuine Zazaki or Gorani verse quoted inside Kurmanji theses.

Limitations

• Questions are machine-written. They were checked only for their answers being in the passage, not read by a person. A question can be broader or narrower than its answer.
• Answers keep the source's errors. They are copied verbatim, so typos, missing capitals and run-together sentences in the source stay (see the example).
• Garbled letters. In kmr_latn, about 1,250 pairs from the theses have letters broken when the PDFs were turned into text, e.g. Ģ for ş ("PiĢtî" for "Piştî"), some theses of classical poetry use symbols such as @, ◊, ¡ or √ for transliterated letters, and many FineWeb-2 pages have þ for ş. kmr_Latn_clean repairs or drops these in the questions and answers, but its contexts can still have them, and garbling not listed in clean.yml remains.
• Near duplicates. 328 pairs share their answer with another pair (338 in kmr_Latn_clean), mostly the same Wikipedia article from ku.wikipedia.org and ku.m.wikipedia.org, or a page FineWeb-2 holds twice. Deduplicate on output before making a test split.
• Topics follow the sources. A third of the pairs come from academic theses (literature, linguistics, history, education) and many web pairs from Wikipedia biographies, so the questions lean towards people, history and literature.

License

The dataset is a derivative of two sources and inherits both licenses:

• Pairs with source HuggingFaceFW/fineweb-2 are under ODC-By 1.0, and also subject to Common Crawl's terms of use.
• Pairs with source muzaffercky/kurdish-theses-yok are under CC BY 4.0.

The questions were generated with Google's Gemini API; check the Gemini API terms for how its outputs may be used.