| # π Misraj Unstructured Data Dump (MUDD) |
|
|
| *A large-scale Arabic text dataset translated from SlimPajama-627B for pretraining Arabic language models* |
|
|
| ## π Dataset Summary |
|
|
| MUDD is a substantial dataset comprising *4,758,338 rows* of unstructured, plain Arabic text. Each entry includes Arabic text and retains its UUID from the original source. This dataset provides high-quality Arabic content specifically designed for pretraining large language models (LLMs) and advancing Arabic natural language processing (NLP) research. |
|
|
| ## π Key Features |
|
|
| * π **Size**: 4,758,338 rows of Arabic text |
| * π£οΈ **Language**: Arabic (translated from English) |
| * π **Source**: Selected subset of [SlimPajama-627B](https://huggingface.co/datasets/cerebras/SlimPajama-627B) |
| * π€ **Translation Model**: [Mutarjim](https://arxiv.org/abs/2505.17894) |
| * π **Format**: Plain text with UUID identifiers |
|
|
| ## ποΈ Dataset Details |
|
|
| ### π Source Data |
|
|
| The foundation of MUDD is [SlimPajama-627B](https://huggingface.co/datasets/cerebras/SlimPajama-627B), a high-quality English text dataset containing: |
|
|
| * π 627 billion tokens |
| * β»οΈ Deduplicated content |
| * π Diverse sources including web pages, Wikipedia, GitHub, and books |
|
|
| *Note*: MUDD is derived from a carefully selected subset of SlimPajama-627B, not the complete dataset. |
|
|
| ### π Translation Process |
|
|
| The Arabic content was generated using *Mutarjim*, a high-performance Arabic-English translation model built on the [Kuwain-1.5B](https://arxiv.org/abs/2504.15120) architecture. The translation involved: |
|
|
| 1. π οΈ *Pre-training*: On extensive monolingual Arabic and English corpora |
| 2. π― *Fine-tuning*: Using high-quality, human-curated parallel sentence pairs for accurate Arabic translations |
|
|
| ## π Dataset Structure |
|
|
| ```json |
| { |
| "uuid": { |
| "dtype": "string", |
| "_type": "Value" |
| }, |
| "plain_text": { |
| "dtype": "string", |
| "_type": "Value" |
| } |
| } |
| ``` |
|
|
| ## π‘ Usage |
|
|
| ### π₯ Loading the Dataset |
|
|
| ```python |
| from datasets import load_dataset |
| |
| dataset = load_dataset("Misraj/mudd") |
| ``` |
|
|
| ### π Example Usage |
|
|
| ```python |
| # Access the first example |
| example = dataset['train'][0] |
| print(f"UUID: {example['uuid']}") |
| print(f"Arabic Text: {example['plain_text']}") |
| ``` |
|
|
| ## π― Intended Use Cases |
|
|
| * π€ **Pretraining Arabic LLMs**: Large-scale, high-quality Arabic text corpus for training new language models |
| * π **Arabic NLP Research**: Supporting research initiatives focused on Arabic language processing |
| * π¦ **Downstream Applications**: Reliable source for projects requiring extensive and diverse Arabic text data |
|
|
| ## π Dataset Statistics |
|
|
| | π Metric | π Value | |
| | ----------------- | ------------------------ | |
| | Total Rows | 4,758,338 | |
| | Language | Arabic | |
| | Source | SlimPajama-627B (subset) | |
| | Translation Model | Mutarjim | |
| | Format | Plain text | |
|
|
| ## π Citations |
|
|
| If you use this dataset, please cite: |
|
|
| ```bibtex |
| @misc{misraj2025mudd, |
| title = {Misraj Unstructured Data Dump (MUDD)}, |
| author = {Khalil Hennara, Muhammad Hreden, Mohamed Motaism Hamed, Zeina Aldallal, Sara Chrouf, Safwan AlModhayan, Ahmed Bustati}, |
| year = {2025}, |
| publisher = {MisrajAI}, |
| howpublished = {\url{[https://huggingface.co/datasets/Misraj/mudd](https://huggingface.co/datasets/Misraj/mudd)}} |
| } |
| ``` |