Datasets:

kurdish-tech
/

kurdish-grammar-eval

	---
	license: cc-by-sa-4.0
	language:
	- kmr
	- ckb
	multilinguality: multilingual
	task_categories:
	- text-classification
	pretty_name: Kurdish Grammar Minimal Pairs (BLiMP-style)
	size_categories:
	- n<1K
	tags:
	- kurdish
	- kurmanji
	- sorani
	- grammar
	- linguistic-acceptability
	- minimal-pairs
	- evaluation
	---

	# Kurdish Grammar Minimal Pairs (BLiMP-style) — Kurmancî · Soranî

	A grammar-competence benchmark for Kurdish, built on the [BLiMP](https://github.com/alexwarstadt/blimp)
	idea: for each item, a correct sentence is paired with a corrupted version where one specific
	grammar rule has been deliberately broken. Score a language model by checking whether it assigns higher
	likelihood to the correct sentence than the corrupted one — accuracy well above 50% means the model
	learned the rule, not just surface fluency.

	Built by [Kurdish-Tech](https://github.com/Kurdish-Tech) on
	[KurdishCorpus-clean](https://huggingface.co/datasets/kurdish-tech/KurdishCorpus-clean).

	## Why this exists

	Perplexity tells you a model sounds fluent. It doesn't tell you whether the model actually learned that
	Kurdish requires an ezafe linker between a noun and its modifier, or that an indefinite marker can't
	stack, or that a copula has to agree with its subject. This benchmark tests exactly that — a narrower,
	harder question than fluency. As far as we've found, there is no other published grammar-competence
	benchmark for Kurdish.

	## What's covered — and what isn't

	Kurmancî and Soranî only. No Zazakî. Every category here required hand-encoding the actual grammar
	rule (the specific suffixes, the specific whitelist of words that can follow them) — this isn't something
	that generalizes automatically from one Kurdish variety to another, and we don't have confident,
	verified knowledge of the equivalent Zazakî rules to encode them correctly. Publishing a guessed Zazakî
	version would risk teaching people the wrong grammar. We'd rather publish two dialects done honestly
	than three with one faked. [Contributions adding Zazakî are genuinely wanted](https://github.com/Kurdish-Tech) —
	open an issue if you have the linguistic background to help.

	## Categories

	\| Category \| Type \| Rule tested \| Kurmancî \| Soranî \|
	\|---\|---\|---:\|---:\|---:\|
	\| `ezafe_deletion` \| auto-mined \| Kurdish noun phrases require an ezafe linker between a head noun and a following possessor/adjective (Kurmancî `-ê/-a/-ên/-yê/-ya/-yên`; Soranî `-ی`). Deleting it is ungrammatical. \| 300 \| 132 \|
	\| `indefinite_duplication` \| auto-mined \| Kurmancî `-ek` / Soranî `-êk` mark indefinite singular and cannot stack. \| 300 \| 270 \|
	\| `copula_agreement` \| hand-authored \| The copula enclitic must agree with subject person/number (Kurmancî `im/î/e/in`; Soranî `-م/-یت/-ه/-ین/-ن`). \| 15 \| 15 \|
	\| Total \| \| \| 615 \| 417 \|

	## Honest limitations — read before quoting a score

	This is a pilot-scale heuristic probe, not a peer-reviewed linguistic benchmark:

	- Scope is narrow on purpose. Three closed-class morphology phenomena, chosen because the rule is
	unambiguous and mechanically checkable — not a general syntax benchmark.
	- Auto-mined items are rule-verified, not item-by-item human-checked. Extraction is restricted to a
	whitelist of common pronouns/adjectives following the candidate suffix specifically to keep precision
	high, but no native speaker read every one of the 1,002 auto-mined pairs individually.
	- `copula_agreement` is small — 15 pairs per dialect. Enough to smoke-test, not enough to trust a
	score difference between two close models. Treat it as a qualitative check, not a leaderboard metric.
	- No Zazakî — see above.
	- Not adversarial. Corruptions are single, clean rule violations, not the harder near-miss
	distractors a fully adversarial benchmark would use.

	## Data provenance & license

	`correct` sentences in the auto-mined categories are drawn verbatim from
	[KurdishCorpus-clean](https://huggingface.co/datasets/kurdish-tech/KurdishCorpus-clean)'s held-out split
	(never seen by the tokenizer or any model trained on the corpus), from openly-licensed sources only:
	CC-100 (Common Crawl ToU), HPLT 2.0 (CC0-1.0), and Kurdish Wikipedia (`ckbwiki`, CC-BY-SA-4.0). The
	`corrupted` counterpart is generated by deleting/duplicating the target morpheme in that same sentence.
	`copula_agreement` items are original, hand-authored by Kurdish-Tech.

	Released under CC BY-SA 4.0.

	## Format

	One JSON object per line, in six files (`{dialect}_{category}.jsonl`):

	```json
	{"id": "cc100:ku:379#ez4", "category": "ezafe_deletion", "correct": "Di nava tirkan da dostên me hebin baş e, ...", "corrupted": "Di nava tirkan da dost me hebin baş e, ..."}
	```

	\| Field \| Description \|
	\|---\|---\|
	\| `id` \| Source document reference for auto-mined items (`gold-{dialect}-copula-N` for hand-authored) \|
	\| `category` \| One of `ezafe_deletion`, `indefinite_duplication`, `copula_agreement` \|
	\| `correct` \| The grammatical sentence \|
	\| `corrupted` \| The same sentence with the target rule violated \|

	## Usage

	```python
	from datasets import load_dataset

	ds = load_dataset("kurdish-tech/kurdish-grammar-eval", data_files="kurmanji_ezafe_deletion.jsonl")

	# Scoring sketch: accuracy = fraction where the model assigns
	# higher likelihood (lower loss) to `correct` than `corrupted`.
	```

	## Citation

	```bibtex
	@misc{kurdishtech2026grammareval,
	title = {Kurdish Grammar Minimal Pairs (BLiMP-style): a pilot grammar-competence benchmark for Kurmanc\^i and Soran\^i},
	author = {{Kurdish-Tech}},
	year = {2026},
	url = {https://huggingface.co/datasets/kurdish-tech/kurdish-grammar-eval}
	}
	```

	---

	*Built by [Kurdish-Tech](https://github.com/Kurdish-Tech) — open-source digital
	infrastructure for the Kurdish language.*

	Maintained by Alan Hesen.