MaLA-LM/mala-opus-dedup-2410

MaLA Corpus: Massive Language Adaptation Corpus

This mala-opus-dedup-2410 is the bilingual part of the MaLA Corpus. It is a cleaned and deduplicated version of OPUS corpus, collected from OPUS with a cutoff of October 2024 (2410). Particularly, it contains bilingual translation data (aka, parallel data or bitexts) in 29,202 language pairs.

The MaLA Corpus (Massive Language Adaptation) is a series of comprehensive, multilingual datasets designed to support the continual pre-training of large language models. This mala-opus-dedup-2410 set can also support the training of multilingual translation models.

Project page: https://mala-lm.github.io

---

Data Fields

source_text: the source text of parallel data.
source_lang: the language of the source text.
target_text: the target text of parallel data.
target_lang: the language of the target text.
original_code - the source-target language codes in the format of ${src_code} - ${trg_code}, delimited by " - ".

---

Key Features

• Language Coverage: Includes data in 29,202 language pairs.

• Pre-processing: The corpus is cleaned and deduplicated to ensure high-quality training data.

---

Dataset Creation

This mala-opus-dedup-2410 set was created by processing data from OPUS, followed by rigorous pre-processing to ensure the quality of the data:

• Cleaning: Noisy and irrelevant data were 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.

We do a variety of handcrafted checks to filter out noisy lines based on this pipeline. In detail, we run line-level script detection to ensure that the writing script identified for the dataset during the preliminary stage makes up more than 5% of that line. We remove lines where the same word or character repeats more than 5 times in a row. We calculate the source and target ratios in the number of characters and number of words as well as alphabetical ratio, but only require one of the ratios to be within our pre-determined range, due to the wide range of scripts, languages, and their usual word delimiters. We also ensure that the line fits between a minimum and a maximum number of character lengths (non-empty).

---

Intended Use

This mala-opus-dedup-2410 set is intended for researchers and developers looking to improve the multilingual capabilities of language models. It is especially useful for:

• Pre-training of large language models, particularly continual pre-training, to enhance the performance in low-resource languages.
• Fine-tuning models on multilingual benchmarks to improve language coverage across a variety of domains.
• Multilingual tasks such as machine translation training or fine-tuning.

---

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 mala-opus-dedup-2410 set was processed by the MaLA-LM project. If you find this dataset useful, please cite our 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}, 
}

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.