wasm_readme.md

## **license: apache-2.0**

# **📚 Misraj Structured Data Dump (MSDD)**

Misraj Structured Data Dump (MSDD) is a large-scale Arabic multimodal dataset created using our **WASM pipeline**. It is extracted and filtered from the [Common Crawl](https://commoncrawl.org/) dumps and uniquely preserves the structural integrity of web content by providing markdown output. This dataset aims to address the lack of high-quality, structured multimodal data for Arabic and accelerate research in large language and multimodal models.


## **📌 Dataset Summary**

- **Source:** Subset from multiple Common Crawl dumps, processed with the WASM pipeline.

- **Documents:** 23 million documents.

- **Timeframe:** \* 2024 Dump: Dump 10

* 2025 Dump: Dump 13

- **Languages:** Primarily Arabic (MSA and dialects).

- **Format:** Multimodal format with interleaved text and images in Markdown.

- **Domain Variety:** General web content.


## **💡 Usage**

### **📥 Loading the Dataset**

from datasets import load\_dataset\
\
dataset = load\_dataset("Misraj/msdd")


### **📋 Example Usage**

\# Access the first example\
example = dataset\['train']\[0]\
print(f"Text: {example\['text']}")\
print(f"Images: {example\['images']}")\
print(f"Captions: {example\['image\_caption']}")


## **⚙️ The WASM Processing Pipeline**

The performance of large language models (LLMs) and large multimodal models (LMMs) depends heavily on the quality and scale of their pre-training datasets. For Arabic, the lack of high-quality multimodal datasets that preserve document structure has limited progress. Our **WASM pipeline** was developed to address this gap by processing Common Crawl and generating a structured, markdown-based multimodal dataset. The pipeline is designed to preserve the structural integrity of web content while maintaining flexibility for both text-only and multimodal pre-training scenarios.

The core of the WASM pipeline involves careful filtering at both the paragraph and document level.


### **✅ _Paragraph-Level Filtering_**

Each paragraph in the corpus undergoes the following checks:

- **Character Deduplication:** Removal of repeated characters beyond a threshold.

- **Word Repetition Ratio:** Filtering paragraphs with excessive word repetitions.

- **Special Character Ratio:** Filtering based on the proportion of non-standard characters.

- **Language Identification:** Only Arabic paragraphs are retained.

- **Perplexity Scoring:** Content scored using an in-house KenLM-based model trained on Wikipedia-like pages, Arabic Twitter data, and dialectal text (e.g., Lahjawi), to remove low-quality text.


### **✅ _Document-Level Filtering_**

Each full document must pass:

- **Word Repetition Ratio:** Similar to paragraph level, but with different thresholds for full documents.

- **Special Character Ratio:** Ensures no document is dominated by symbols, code snippets, or garbage text.

- **Language Identification:** Verifies the document is primarily Arabic.

- **Perplexity Score:** Documents are filtered based on perplexity thresholds to maintain fluent, natural text.


## **📂 Dataset Structure**

The dataset is structured with three main columns to support multimodal tasks. The text is interleaved with image placeholders, allowing for rich text-and-image documents.

- **text**: A string containing the textual content. The special token \<image> is used to denote the position where an image should be inserted.

- **images**: A list of image URLs (strings). These images correspond sequentially to the \<image> tokens in the text field.

- **image\_caption**: A list of strings, where each string is a caption for the corresponding image in the images list. If an image does not have a caption, the list will contain an empty string '' at that position.

The dataset has the following features:
```json
{\
  "text": {\
    "dtype": "string",\
    "\_type": "Value"\
  },\
  "images": {\
    "feature": {\
      "dtype": "list",\
      "\_type": "Value"\
    },\
    "\_type": "Sequence"\
  },\
  "image\_caption": {\
    "feature": {\
      "dtype": "list",\
      "\_type": "Value"\
    },\
    "\_type": "Sequence"\
  }\
}
```

## **🚦 Quality Checks**

The dataset quality was validated using:

- In-house KenLM-based Arabic models for perplexity checks.

- Manual inspection of samples.

- A pipeline inspired by [OBELICS](https://github.com/huggingface/OBELICS), with custom enhancements.

- Comparative analysis against major existing dataset processing pipelines to validate design choices.


## **🔍 Intended Use**

This dataset is intended for:

- Training large-scale multimodal Arabic language models.

- Research on Arabic NLP, including dialect modeling and low-resource language studies.


## **🌐 Availability & Reproducibility**

To support future research and ensure reproducibility, we are publicly releasing this representative dataset dump. The WASM processing pipeline for Arabic will also be made available to the community.


## **📝 Citation**

If you use this dataset, please cite:
```bibtex
@misc{misraj2025msdd,\
  title        = {Misraj Structured Data Dump (MSDD)},\
  author       = {Khalil Hennara, Muhammad Hreden, Mohamed Motasim Hamed, Zeina Aldallal, Sara Chrouf, Safwan AlModhayan, Ahmed Bustati},\
  year         = {2025},\
  publisher    = {MisrajAI},\
  howpublished = {\url{\[https\://huggingface.co/datasets/Misraj/msdd]\(https\://huggingface.co/datasets/Misraj/msdd)}}\
}
```