shkomq/Kurdish_Multi-Domain_Corpus_KMDC

Kurdish Multi-Domain Corpus (KMDC)

Dataset Description

The Kurdish Multi-Domain Corpus (KMDC) is a large-scale instruction-style dataset designed to support natural language processing (NLP), supervised fine-tuning (SFT), and large language model (LLM) development for Central Kurdish (Sorani). The dataset consists of structured question–response pairs generated through an LLM-guided pipeline that transforms raw Kurdish text into machine-learning-ready conversational data across multiple domains.

KMDC aims to address the scarcity of high-quality Kurdish datasets for generative AI and low-resource language research. The corpus covers diverse categories such as news, healthcare, sports, history, education, information technology, culture, and general knowledge, making it suitable for a wide range of downstream NLP applications.

Dataset Statistics

Metric	Value
Dataset Name	Kurdish Multi-Domain Corpus (KMDC)
Total Entries	133,115
Total Word Count	~5,084,425
Estimated Tokens	~7,626,637
File Format	JSON
File Size	74.96 MB
Language	Central Kurdish (Sorani)
Language Code	ckb
Region	Kurdistan-Iraq
Dataset Type	Instruction-style Question–Response Pairs
Domains	Multi-Domain
Primary Use	Instruction Tuning and Supervised Fine-Tuning (SFT)

Dataset Structure

The dataset is provided in JSON format. Each entry follows the structure below:

{
  "id": "CKB-KMDC-[HASH]",
  "category": "String representing the domain or category",
  "question": "The question in Central Kurdish (Sorani)",
  "response": "The corresponding response in Central Kurdish (Sorani)"
}

Languages

The primary language of this dataset is Central Kurdish (Sorani) (ckb), specifically focusing on the dialect used in the Kurdistan Region of Iraq.

Dataset Construction

KMDC was constructed using a multi-stage pipeline involving:

1. Collection of raw Kurdish textual data from multiple domains
2. Text preprocessing and normalization
3. LLM-guided transformation into structured question–response pairs
4. Quality filtering and validation

This pipeline enables scalable creation of structured datasets for low-resource language model development.

Use Cases

• Question Answering (QA)
• Instruction Fine-Tuning for Large Language Models (LLMs)
• Conversational AI and Chatbots
• Retrieval-Augmented Generation (RAG)
• General Kurdish Text Generation
• Low-Resource Language NLP Research

Limitations

Although KMDC provides broad multi-domain coverage, some samples may contain automatically generated inconsistencies, domain imbalance, or contextual limitations. Additional human validation and filtering are recommended for high-stakes applications.

Citation

@dataset{kmdc2026,
  title={Kurdish Multi-Domain Corpus (KMDC)},
  author={Shko Muhammed Qader and Zheng Wang},
  year={2026},
  publisher={Hugging Face}
}