yousef-khoubrane/tulu-v2-sft-mixture-english-darija

TÜLU-v2-SFT-Mixture-English-Darija

The TÜLU-v2-SFT-Mixture-English-Darija dataset aligns English instructions from AllenAI's TÜLU-v2-SFT-Mixture with their Moroccan Darija translations from MBZUAI-Paris's Darija-SFT-Mixture. It supports tasks like machine translation, instruction-following, and cross-lingual understanding.

Dataset Details

• English Instructions: Sourced from AllenAI's TÜLU-v2-SFT-Mixture, a mixture of diverse instruction-tuning datasets.
• Darija Translations: From MBZUAI-Paris's Darija-SFT-Mixture, an instruction-tuning dataset for Darija LLMs that combines instructions from various sources. I filtered the subsets corresponding to the Darija translations of TÜLU-v2 based on their IDs.

Usage

This dataset is ideal for training and fine-tuning language models in English and Moroccan Darija, enhancing accessibility for Moroccan Arabic speakers.

Notes

• Not all original English samples have Darija translations. This is because some filtering was done both before and after the translation process to ensure the quality and relevance of Darija-SFT-Mixture. This dataset only includes rows with existing translations.
• Rows with duplicated IDs were excluded from this version of the dataset. In the original English dataset, some IDs were repeated, each corresponding to a slightly different variation of the same instruction. These variations were then translated into Darija. The challenge here is that the usual method of mapping by ID doesn't work, as multiple translations exist for a single ID, and it's difficult to determine which translation matches which variation. This issue needs to be resolved before including these rows in the dataset.

Licensing

The dataset is released under the ODC-BY license. Some portions are non-commercial; review licenses for each subset to ensure compliance.

Citation

If you use this dataset, please cite the original sources:

@misc{ivison2023camels,
      title = {Camels in a Changing Climate: Enhancing LM Adaptation with Tulu 2},
      author = {Ivison, Hamish and Wang, Yizhong and Pyatkin, Valentina and Lambert, Nathan and Peters, Matthew and Dasigi, Pradeep and Jang, Joel and Wadden, David and Smith, Noah A. and Beltagy, Iz and Hajishirzi, Hannaneh},
      year = {2023},
      eprint = {2311.10702},
      archivePrefix = {arXiv},
      primaryClass = {cs.CL}
}
@article{shang2024atlaschat,
      title = {Atlas-Chat: Adapting Large Language Models for Low-Resource Moroccan Arabic Dialect},
      author = {Shang, Guokan and Abdine, Hadi and Khoubrane, Yousef and Mohamed, Amr and Abbahaddou, Yassine and Ennadir, Sofiane and Momayiz, Imane and Ren, Xuguang and Moulines, Eric and Nakov, Preslav and Vazirgiannis, Michalis and Xing, Eric},
      year = {2024},
      eprint = {2409.17912},
      archivePrefix = {arXiv},
      primaryClass = {cs.CL},
      url = {https://arxiv.org/abs/2409.17912}
}

---