README.md

---
language:
- ku
- ckb
- kmr
- hac
size_categories:
- 1B<n<10B
license: cc-by-4.0
pretty_name: KurCorpus 2B
tags:
- kurdish
- low-resource
- language-modeling
---

# KurCorpus 2B

![License: CC BY 4.0](https://img.shields.io/badge/License-CC%20BY%204.0-blue)
![Language: Kurdish (ckb/kmr/hac)](https://img.shields.io/badge/Language-Kurdish%20(ckb%2Fkmr%2Fhac)-informational)
![Size: 2B+ tokens](https://img.shields.io/badge/Size-2B%2B%20tokens-brightgreen)
[![DOI](https://img.shields.io/badge/DOI-10.17632%2Ffb5xhhn6m5.1-orange)](https://doi.org/10.17632/fb5xhhn6m5.1)

**KurCorpus 2B** is a multidialectal Kurdish text corpus (>2B tokens) for large-scale language modeling and downstream NLP.

- **Dialects:** Sorani (ckb), Kurmanji/Badini (kmr), Hawrami/Gorani (hac)  
- **License:** CC BY 4.0  
- **Repo:** https://huggingface.co/datasets/abdulhade/Kurdishcorpus  
- **External record:** Mendeley Data DOI `10.17632/fb5xhhn6m5.1`

---

## TL;DR

- Ready for **pretraining** and **finetuning** Kurdish LMs  
- Single field **`text`** (UTF-8), offered as large archives or sharded `.txt(.gz)`  
- Includes **normalization and cleaning** (Unicode, orthography, noise removal with placeholders like `[URL]`, `[EMAIL]`)  
- No official splits; create your own task-specific splits and report dialect coverage

---

## Quickstart

```python
from datasets import load_dataset

# Stream without downloading everything at once (recommended for very large corpora)
ds = load_dataset("abdulhade/Kurdishcorpus", split="train", streaming=True)

for i, ex in enumerate(ds.take(5)):
    print(ex["text"])
```
# Text Normalization (AsoSoft Library)

This repository documents how Central Kurdish text is normalized with the **AsoSoft Library** before release. The goal is to standardize Unicode, punctuation, numerals, and common web artifacts so downstream tokenizers and language models see clean, consistent text.

- Library: **AsoSoft Library** (C#/.NET) by Aso Mahmudi  
- Repo: https://github.com/AsoSoft/AsoSoft-Library  
- Scope: Kurdish Sorani (ckb) primary, with script-agnostic steps for other Kurdish varieties (kmr, hac)

---

## Why normalize

- Reduce spurious vocabulary caused by look-alike glyphs and legacy encodings.
- Ensure consistent numerals and punctuation for stable tokenization.
- Replace privacy-sensitive artifacts like URLs and emails with safe placeholders.
- Remove invisible noise (tatweel, zero-widths, control chars) that hurts training.

---

## What the pipeline does

### 1) Legacy encodings to Unicode (conditional)
Applied only when legacy font artifacts are detected.

- `AliK2Unicode`
- `AliWeb2Unicode`
- `Dylan2Unicode`
- `Zarnegar2Unicode`

### 2) HTML and web artifacts
- `ReplaceHtmlEntity` converts entities (e.g. `&quot;` → `"`).
- `ReplaceUrlEmail` replaces URLs and emails with `[URL]` and `[EMAIL]`.

### 3) Unicode standardization and Kurdish-specific fixes
`Normalize(...)` with deep and additional corrections enabled.

- Replace Arabic Presentation Forms with standard code points.
- Remove Tatweel U+0640 and stray control/zero-width characters.
- Normalize dashes and spaces.
- Map visually similar Arabic letters to Kurdish standard forms (e.g. ڪ, ے, ٶ → ک, ی, ؤ).
- Kurdish-specific standardization of `ە, هـ, ی, ک`.
- Correct frequent mis-conversions from non-Unicode fonts.
- Optionally convert word-initial `ر` to `ڕ`.

### 4) Punctuation and whitespace
- `NormalizePunctuations(text, separateAllPunctuations=false)` ensures no extra spaces around punctuation.
- `TrimLine` trims leading/trailing whitespace including zero-width spaces.

### 5) Numerals and digit separation
- `UnifyNumerals(text, "en")` normalizes all digits to ASCII 0–9 for tokenizer stability.
- `SeperateDigits` inserts spaces between digits and adjacent letters (e.g. `12کەس` → `12 کەس`).

> Note: For kmr/hac lines, we only apply script-agnostic steps (HTML/entity cleanup, URL/email replacement, Unicode cleanup, punctuation, numerals, trimming). We do not transliterate or run G2P on released text.

---

## Reproducible configuration

We run the following configuration in order. The last step that applies is the one whose conditions are met.

```csharp
using AsoSoftLibrary;

string NormalizeCkbLine(string text)
{
    // 1) Legacy converters (conditional, heuristic triggers)
    if (LooksLikeAliK(text))      text = AsoSoft.AliK2Unicode(text);
    if (LooksLikeAliWeb(text))    text = AsoSoft.AliWeb2Unicode(text);
    if (LooksLikeDylan(text))     text = AsoSoft.Dylan2Unicode(text);
    if (LooksLikeZarnegar(text))  text = AsoSoft.Zarnegar2Unicode(text);

    // 2) HTML and web artifacts
    text = AsoSoft.ReplaceHtmlEntity(text);
    text = AsoSoft.ReplaceUrlEmail(text); // replaces with [URL], [EMAIL]

    // 3) Unicode standardization (+ Kurdish fixes)
    // changeInitialR=true enables initial ر → ڕ
    text = AsoSoft.Normalize(
        text: text,
        isOnlyKurdish: true,
        changeInitialR: true,
        deepUnicodeCorrectios: true,
        additionalUnicodeCorrections: true,
        usersReplaceList: null
    );

    // 4) Punctuation and whitespace
    text = AsoSoft.NormalizePunctuations(text, separateAllPunctuations: false);
    text = AsoSoft.TrimLine(text);

    // 5) Numerals and digit separation
    text = AsoSoft.UnifyNumerals(text, "en");
    text = AsoSoft.SeperateDigits(text);

    return text;
}
```

Heuristic detectors like `LooksLikeAliK` check for characteristic code points and glyph ranges associated with each legacy font family. The converters are applied only when such patterns are present.

---

## Before and after examples

```
Input : دەقی«کوردی » و ڕێنووس ،((خاڵبەندی )) چۆنە ؟
Output: دەقی «کوردی» و ڕێنووس، «خاڵبەندی» چۆنە?
```

```
Input : ژمارەکانی ٤٥٦ و ۴۵۶ و 456
Output: ژمارەکانی 456 و 456 و 456
```

```
Input : دەقے شیَعري خـــۆش. ره‌نگه‌كاني خاك
Output: دەقی شێعری خۆش. ڕەنگەکانی خاک
```

```
Input : ئێوە &quot;دەق&quot; لە زمانی &lt;کوردی&gt; دەنووسن
Output: ئێوە "دەق" بە زمانی <کوردی> دەنووسن
```

```
Input : لە ساڵی1950دا1000دۆلاریان بە 5کەس دا
Output: لە ساڵی 1950 دا 1000 دۆلاریان بە 5 کەس دا
```

---

## What we do not do

- No transliteration or G2P on the released corpus. The text remains in its native script.
- No poem meter classification during release processing.
- No semantic rewriting beyond canonical character and punctuation normalization.

---

## Dialect coverage policy

- **ckb**: full AsoSoft normalization including Kurdish-specific fixes.
- **kmr/hac**: script-agnostic cleanup only. We avoid ckb-specific letter substitutions to preserve dialectal orthography and phonology.

---

## Quality checks

- Character inventory auditing before and after normalization.
- Proportion of placeholders `[URL]`, `[EMAIL]` to detect crawler bias.
- Count of removed control and zero-width characters.
- Round-trip sampling with spot manual verification on ambiguous lines.
- Determinism check: same input produces same output across runs.

---

## Reproducing locally

1. Install .NET 6 or newer.
2. Add AsoSoft Library via NuGet in your C# project.
3. Pipe your text through a console app using the configuration above.

Example console usage:

```bash
dotnet run --project NormalizerApp --input data/raw.txt --output data/normalized.txt
```

Where `NormalizerApp` calls `NormalizeCkbLine` for each line.

---

## Parameters you can tune

- `changeInitialR`: set to `false` if you do not want `ر` to become `ڕ` word-initially.
- `UnifyNumerals` target: `"en"` for ASCII 0–9 or `"ar"` for Arabic-Indic digits, depending on your tokenizer.
- `separateAllPunctuations`: set to `true` if you prefer all punctuation to be isolated by spaces for whitespace-tokenized pipelines.

---

## Cite

If you use **KurCorpus 2B**, please cite:

[![DOI](https://img.shields.io/badge/DOI-10.17632%2Ffb5xhhn6m5.1-orange)](https://doi.org/10.17632/fb5xhhn6m5.1)

**BibTeX**
```bibtex
@dataset{rawf2025kurcorpus2b,
  title   = {KurCorpus 2B: A Multidialectal 2-Billion-Token Corpus for Kurdish Language Modeling},
  author  = {Rawf, Karwan Mahdi and Abdullah, Abdullah and Hussein, Amanj and Mohammed, Haukar},
  year    = {2025},
  version = {1},
  howpublished = {Mendeley Data},
  doi     = {10.17632/fb5xhhn6m5.1}
}