premio-ai/TheArabicPile_Web

The Arabic Pile

Introduction:

The Arabic Pile is a comprehensive dataset meticulously designed to parallel the structure of The Pile and The Nordic Pile. Focused on the Arabic language, the dataset encompasses a vast array of linguistic nuances, incorporating both Modern Standard Arabic (MSA) and various Levantine, North African, and Egyptian dialects. Tailored for the training and fine-tuning of large language models, the dataset consists of 13 subsets, each uniquely crafted to cater to different linguistic domains.

The Web Subset:

This dataset has a collection of Arabic web articles based on Common Crawl.

Other Subsets:

1. premio-ai/TheArabicPile
2. premio-ai/TheArabicPile_Web
3. premio-ai/TheArabicPile_Lyrics
4. premio-ai/TheArabicPile_Reviews
5. premio-ai/TheArabicPile_Dialects
6. premio-ai/TheArabicPile_Mathematics
7. premio-ai/TheArabicPile_Conversational
8. premio-ai/TheArabicPile_Articles
9. premio-ai/TheArabicPile_Poetry
10. premio-ai/TheArabicPile_Medical
11. premio-ai/TheArabicPile_Miscellaneous
12. premio-ai/TheArabicPile_SocialMedia
13. premio-ai/TheArabicPile_Translations
14. premio-ai/TheArabicPile_Books

These subsets serve distinct purposes, ranging from mathematical content to conversational dialogue, medical texts, and more. Notably, there's a dedicated subset, "premio-ai/TheArabicPile_SocialMedia," emphasizing the inclusion of language commonly found in social media contexts.

Data Structure

The datasets are divided into two main subsets:

1. Original Subset: The raw data as collected from sources, without modifications.
2. Deduplication Subset: A filtered and cleaned version, enhancing usability for large language models by reducing redundancy and noise. The Arabic Pile extends an invitation not only for training and fine-tuning large language models but also for diverse applications across linguistic domains. Whether for research, analysis, or other linguistic endeavors, The Arabic Pile stands as a rich resource for the exploration of Arabic language intricacies.

Data Collection

Please refer to the paper for more details on our data collection procedures.

Data Format

The dataset has one single column called text. The text should contain the required meta data and the body combined. This was done to make sure that it will be a good fit for direct training or fine-tuning of large language models.

Please note that the meta data might require to be repeated if your training context window won’t fit the entire body of text.

Potential Bias

As with any large-scale dataset, The Arabic Pile is not immune to potential biases that may influence the training and performance of language models. It's crucial to transparently address these biases to ensure responsible usage and interpretation of the dataset. Here are some considerations:

1. Dialectal Imbalance: The dataset incorporates various Arabic dialects, with a focus on Levantine, North African, and Egyptian variants. However, there might be variations in the representation of these dialects, potentially leading to an imbalance in the training data.
2. Source Influence: Bias may arise from the sources of the original data. The dataset collects information from diverse platforms and domains, and biases inherent in those sources could transfer to the dataset.
3. Social Media Context: Some of our datasets have language from social media platforms and online platforms. This subset may introduce biases inherent in online discourse, such as informal language, colloquial expressions, and potential subjectivity in politics, religion or culture.
4. Genre and Domain Bias: Different subsets cater to distinct linguistic domains, such as medical texts, poetry, reviews, and more. Each domain carries its own linguistic characteristics, potentially leading to biases based on the genres represented.

License Information for The Arabic Pile: No Commercial Use

The Arabic Pile is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). This license is designed to facilitate the open sharing and collaboration of the dataset while ensuring responsible and non-commercial usage. Key Points of the License:

• Attribution (BY): Users are free to share, adapt, and build upon the dataset, even commercially, as long as they provide appropriate attribution to the dataset creators.
• Non-Commercial (NC): The dataset may not be used for commercial purposes. Any use for commercial gain requires explicit permission from the dataset creators.
• No Additional Restrictions: The license allows for maximum freedom of use, provided the terms of attribution and non-commercial use are adhered to. How to Cite: When using The Arabic Pile in your work, please include a proper citation to acknowledge the dataset creators. A recommended citation can be found in the model card for easy reference. License Deed: For a comprehensive understanding of the terms and conditions, please refer to the CC BY-NC 4.0 License Deed. By adopting this license, we aim to foster a collaborative and open environment for the exploration and advancement of Arabic language understanding and natural language processing.

Citation

When utilizing The Arabic Pile in your research, development, or other projects, we kindly request that you cite the dataset using the following format:

@article{alrefaie2024arabicpile, author = {Mohamed Taher Alrefaie, Mahmoud Ibrahim Barbary, Ahmed Yasser Hassanein, Shiref Khaled Elhalawany, Karim Ashraf Elsayed, Ahmed Yasser }, title = {The Arabic Pile: A Large Scale Dataset of Diverse Text for Large Language Modeling}, year = {2024}, url = {https://huggingface.co/datasets/premio-ai/TheArabicPile} }