Update README.md

README.md

@@ -85,4 +85,115 @@ configs:
   data_files:
   - split: 2026_06_27_part
     path: kmr_Latn/2026_06_27_part-*
+language:
+- ku
+- kmr
+- ckb
+- diq
+license: cc-by-4.0
+pretty_name: Kurdish Web Corpus
+size_categories:
+- 100K<n<1M
+task_categories:
+- text-generation
+- fill-mask
 ---
+
+# Dataset Card for Kurdish Web Corpus (Deduplicated)
+
+## Dataset Summary
+
+A multilingual corpus of web text in three Kurdish/Zaza varieties — Kurmanji Kurdish
+(`kmr_Latn`), Sorani Kurdish (`ckb_Arab`), and Zazaki (`diq_Latn`) — scraped from Kurdish-language
+websites, language-identified with [GlotLID](https://github.com/cisnlp/GlotLID), and
+deduplicated (exact + MinHash near-duplicate removal). See the "Dataset Creation" section below
+for the full pipeline.
+
+Rows, by language config:
+
+| config      | language          | script | rows    | publishers |
+|-------------|-------------------|--------|---------|------------|
+| `kmr_Latn`  | Kurmanji Kurdish  | Latin  | 517,335 | 51         |
+| `ckb_Arab`  | Sorani Kurdish    | Arabic | 158,266 | 15         |
+| `diq_Latn`  | Zazaki            | Latin  | 319     | 12         |
+
+## Supported Tasks
+
+Intended as pretraining/fine-tuning text for causal or masked language modeling in
+under-resourced Kurdish and Zaza varieties. No task-specific labels are provided.
+
+## Languages
+
+- `kmr` — Kurmanji (Northern Kurdish), Latin script
+- `ckb` — Sorani (Central Kurdish), Arabic script
+- `diq` — Zazaki (Dimli), Latin script
+
+Language and `lang_score` were assigned by GlotLID at scrape time (see
+[kurdish_scrapy](https://github.com/cikay/kurdish_scrapy)) and are carried through unchanged;
+they are not re-verified by the dedup pipeline.
+
+## Dataset Structure
+
+### Data Instances
+
+```json
+{
+  "url": "https://example.com/page-slug",
+  "text": "Full page text ...",
+  "lang": "kmr_Latn",
+  "lang_score": 0.98,
+  "publish_date": "2024-01-01",
+  "publisher": "example.com",
+  "title": "Page title",
+  "word_count": 342
+}
+```
+
+### Data Fields
+
+- `url` (`string`): the page's source URL; doubles as a unique document ID.
+- `text` (`string`): full page text, NFC-normalized.
+- `lang` (`string`): GlotLID language/script code, one of `kmr_Latn`, `ckb_Arab`, `diq_Latn`.
+- `lang_score` (`float32`): GlotLID confidence score for `lang`.
+- `publish_date` (`string`, nullable): publication/posting date as recorded by the source site.
+- `publisher` (`string`): source site's domain.
+- `title` (`string`): page title.
+- `word_count` (`int32`): whitespace-split word count of `text`, computed from the final
+  deduplicated text (not carried over from the source).
+
+### Data Splits
+
+There is no train/dev/test split. Instead, the split name is a crawl snapshot date: it states
+that all data was collected up to and including that date. For example, a split named
+`2026_6_27` means the crawl covers everything gathered through June 27, 2026.
+
+A `_part` suffix (e.g. `2026_6_27_part`) means the crawl was interrupted before it finished —
+it did **not** manage to collect all data through that date, so treat it as a partial/incomplete
+snapshot rather than a clean cutoff. Later, complete uploads may supersede an earlier `_part`
+snapshot for the same or a later date.
+
+Row counts for the current snapshot:
+
+| config      | rows    |
+|-------------|---------|
+| `kmr_Latn`  | 517,335 |
+| `ckb_Arab`  | 158,266 |
+| `diq_Latn`  | 319     |
+
+## Dataset Creation
+
+### Source Data
+
+Pages were scraped from Kurdish-language websites with
+[kurdish_scrapy](https://github.com/cikay/kurdish_scrapy), with page text extracted via
+[trafilatura](https://github.com/adbar/trafilatura) and language-identified with GlotLID at
+scrape time. Exact- and near-duplicate removal was then run with
+[dedup](https://github.com/cikay/dedup) (this repo).
+
+## Considerations for Using the Data
+
+- Publisher representation is highly uneven, especially for `kmr_Latn` and `ckb_Arab`, where a
+  single site (`anf-news.com`, `kurdistan24.net`) accounts for a large share of rows — expect
+  the corresponding editorial style/bias to dominate.
+- `diq_Latn` is extremely small (319 rows) relative to the other two varieties.
+- `lang`/`lang_score` reflect automatic language ID and were not manually verified.