Markhor-Data

Markhor-Data contains the datasets used in:

Markhor: A Framework for Adapting LLMs to Urdu via Alignment-Free Tokenizer Replacement
Findings of the Association for Computational Linguistics: AACL-IJCNLP 2026.

The repository provides the data used for continual pretraining, instruction tuning, and downstream evaluation of Urdu language models developed as part of the Markhor framework.

Dataset Structure

Markhor-Data/
├── pretraining/
│   └── urdu_train_text_new.txt
├── instruction_tuning/
│   ├── urdu_instructions.json
│   ├── urdu_instructions_gpt4o.json
│   └── all_urdu_instructions_deepsk_gpt4o.json
└── evaluation/
    ├── classification/
    │   ├── emotion.xlsx
    │   ├── fnd1.xlsx
    │   ├── fnd2.xlsx
    │   ├── fnd3.xlsx
    │   └── hate_speech.xlsx
    └── qa/
        ├── indomain/
        │   ├── test_biology.json
        │   ├── test_chemistry.json
        │   ├── test_geography.json
        │   ├── test_history.json
        │   └── test_physics.json
        └── outofdomain/
            ├── biology_qa_urdu.json
            ├── chemistry_qa_urdu.json
            ├── geography_qa_urdu.json
            ├── history_qa_urdu.json
            └── physics_qa_urdu.json

Pretraining Corpus

The continual-pretraining corpus contains Urdu text compiled from multiple sources, including Urdu news articles, Wikipedia, and Urdu books.

The corpus includes material from the publicly available Urdu News Dataset by Saurabh Shahane:

https://www.kaggle.com/datasets/saurabhshahane/urdu-news-dataset

The Urdu News Dataset is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Additional Urdu news material, Wikipedia text, and resources provided by the Center for Language Engineering (CLE) were incorporated during corpus construction. The collected text was cleaned and normalized before being used for continual pretraining.

The released pretraining file represents the combined corpus used by the Markhor training pipeline.

Instruction-Tuning Data

The instruction-tuning collection contains 103,463 Urdu instruction-response pairs generated using two teacher models:

• GPT-4o-mini
• DeepSeek

The repository contains the teacher-specific datasets as well as the combined dataset used for instruction tuning.

The files are:

• urdu_instructions.json — instruction-response pairs generated using DeepSeek.
• urdu_instructions_gpt4o.json — instruction-response pairs generated using GPT-4o-mini.
• all_urdu_instructions_deepsk_gpt4o.json — combined instruction-tuning dataset.

The instruction data cover a broad range of topics and were developed to provide Urdu instruction-following and general-knowledge supervision.

Evaluation Data

Classification

The evaluation/classification/ directory contains the datasets used to evaluate Markhor models on Urdu downstream classification tasks.

The evaluated tasks include:

• Fake news detection
• Hate speech detection
• Emotion classification

The fake news evaluation consists of three datasets (fnd1, fnd2, and fnd3). All classification datasets use the columns text and label, with zero-based integer class labels.

These classification datasets originate from previously published Urdu NLP resources. Full citations for the original datasets are provided in the Markhor paper. For details on the class labels and their definitions, users should refer to the corresponding original publications and dataset sources. Users should also consult the original sources for dataset-specific attribution and usage conditions.

Question Answering

The Urdu QA benchmark evaluates generation across five knowledge domains:

• Biology
• Chemistry
• Geography
• History
• Physics

Two evaluation settings are provided:

• In-domain — held-out questions corresponding to domains represented during knowledge elicitation.
• Out-of-domain — independently prepared Urdu question-answer pairs used to evaluate generalization beyond the knowledge-elicitation data.

Each domain contains 100 question-answer pairs.

Download

The complete dataset repository can be downloaded using the Hugging Face CLI:

hf download ma1993/Markhor-Data \
    --repo-type dataset \
    --local-dir data

This creates the data/ directory structure expected by the official Markhor implementation.

Code

The official implementation of Markhor is available at:

https://github.com/zainali93/Markhor

Intended Use

Markhor-Data is released to support research on Urdu natural language processing, low-resource language modeling, tokenizer adaptation, continual pretraining, instruction tuning, and language-model evaluation.

Users should consider the provenance and applicable terms of individual source resources when redistributing or reusing their contents.

Citation

If you use Markhor-Data or the resources released with Markhor, please cite:

@inproceedings{ali2026markhor,
  title     = {Markhor: A Framework for Adapting LLMs to Urdu via Alignment-Free Tokenizer Replacement},
  author    = {Ali, Muhammad Zain and Wang, Yuxia and Manzoor, Muhammad Arslan and Smith, Tony and Pfahringer, Bernhard},
  booktitle = {Findings of the Association for Computational Linguistics: AACL-IJCNLP 2026},
  year      = {2026}
}

The citation will be updated with the official ACL Anthology entry after publication.

Related Resources

• Code: https://github.com/zainali93/Markhor
• Dataset: https://huggingface.co/datasets/ma1993/Markhor-Data

Disclaimer

Some components of Markhor-Data incorporate or derive from third-party resources. Users are responsible for complying with the licenses, terms of use, and attribution requirements associated with the respective original sources.