MaLA-LM/mala-monolingual-integration

MaLA Corpus: Massive Language Adaptation Corpus

This is the noisy version that integrates texts from different sources.

Dataset Summary

The MaLA Corpus (Massive Language Adaptation) is a comprehensive, multilingual dataset designed to support the continual pre-training of large language models. It covers 939 languages and consists of over 74 billion tokens, making it one of the largest datasets of its kind. With a focus on improving the representation of low-resource languages, the MaLA Corpus is a critical resource for advancing multilingual models, particularly those aimed at serving underrepresented languages.

• Project page: https://mala-lm.github.io/emma-500
• Paper: https://arxiv.org/abs/2409.17892

---

Key Features

• Language Coverage: Includes data for 939 languages, with 546 languages having over 100,000 tokens.
• Pre-processing: The corpus is cleaned and deduplicated to ensure high-quality training data.

---

Dataset Structure

The MaLA Corpus is structured to accommodate a wide variety of data types and tasks:

• Languages: The dataset spans 939 languages. The top 546 languages have over 100k tokens, with the remaining 393 languages contributing smaller but valuable amounts of data.
• Tokens: More than 74 billion tokens in total, making it suitable for training large multilingual models.

Dataset Creation

The MaLA Corpus was created by aggregating data from a variety of sources, followed by rigorous pre-processing to ensure the quality of the data:

• Cleaning: Noisy and irrelevant data was removed to ensure higher data quality.
• Deduplication: Duplicate entries across multiple sources were eliminated.
• Normalization: The data was normalized, and language codes were standardized to ISO 639-3 to ensure consistency across all sources.

---

Intended Use

The MaLA Corpus is intended for researchers and developers looking to improve the multilingual capabilities of language models. It is especially useful for:

• Continual pre-training of large language models, such as Llama or XLM-R, to enhance their performance in low-resource languages.
• Multilingual tasks such as machine translation, open-ended generation, and commonsense reasoning.
• Training and fine-tuning models on multilingual benchmarks to improve language coverage across a variety of domains.

---

Take-down Policy

We don't own any part of the data. The original source of each data point is indicated in the collection and source fields.
We will comply with legitimate requests by removing the affected sources from the corpora.

---

Citation

This dataset is compiled and released in the paper below.

@article{ji2024emma500enhancingmassivelymultilingual,
      title={{EMMA}-500: Enhancing Massively Multilingual Adaptation of Large Language Models}, 
      author={Shaoxiong Ji and Zihao Li and Indraneil Paul and Jaakko Paavola and Peiqin Lin and Pinzhen Chen and Dayyán O'Brien and Hengyu Luo and Hinrich Schütze and Jörg Tiedemann and Barry Haddow},
      year={2024},
      journal={arXiv preprint 2409.17892},
      url={https://arxiv.org/abs/2409.17892}, 
}

The final version of this dataset 🤗MaLA-LM/mala-monolingual-split is also used for training the models presented in the below paper

@article{ji2025emma2,
      title={Massively Multilingual Adaptation of Large Language Models Using Bilingual Translation Data}, 
      author={Shaoxiong Ji and Zihao Li and Jaakko Paavola and Indraneil Paul and Hengyu Luo and Jörg Tiedemann},
      year={2025},
      journal={arXiv preprint 2506.00469},
      url={https://arxiv.org/abs/2506.00469},
}

Acknowledgements

We extend our thanks to the language communities and contributors who helped source, clean, and validate the diverse data used in the MaLA Corpus. Their efforts are invaluable in supporting linguistic diversity in AI research.

This work is done by researchers at Helsinki-NLP in collaboration with partners from TU Darmstadt, the University of Edinburgh, and LMU Munich. It is funded by HPLT and UTTER.