Dataset Card for xMMEB-train

Dataset Description

This is a translated version of MMEB-train obtained via machine translation using the MADLAD 3B model. The dataset has been translated to 4 languages: French, German, Italian and Spanish. Please note that we used a specific version of the MMEB dataset where 26 tasks were available (rather than 21). The commit id for the version that we used is the following: "0c3f4b8".

Files are formatted following this standard:

MMEB_{dataset_name}_train_{lang}.jsonl

Where:

• dataset_name: is the original dataset used (e.g. "A-OKVQA");
• lang: is the language of the dataset and task pair (one of "de", "fr", "en", "es", "it").

Note that each file contains at most the first 10.000 instances of the original dataset (we fix this threshold since there are datasets with less overall number of instances).

Additionally, we release a "parallel_shuffled.jsonl" file. It has been formatted to match the formatting of the original dataset and prepared for parallel corpus training. Therefore, each instance of the file contains an instruction in a non-English language and its translation to English.

For the images, you can download them from the original dataset.

If you use this dataset, you should cite the original work that introduced the dataset and ours as well. Specifically:

xMMEB

@misc{musacchio2025xvlm2vecadaptinglvlmbasedembedding,
      title={xVLM2Vec: Adapting LVLM-based embedding models to multilinguality using Self-Knowledge Distillation}, 
      author={Elio Musacchio and Lucia Siciliani and Pierpaolo Basile and Giovanni Semeraro},
      year={2025},
      eprint={2503.09313},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.09313}, 
}

MMEB

@article{jiang2024vlm2vec,
  title={Vlm2vec: Training vision-language models for massive multimodal embedding tasks},
  author={Jiang, Ziyan and Meng, Rui and Yang, Xinyi and Yavuz, Semih and Zhou, Yingbo and Chen, Wenhu},
  journal={arXiv preprint arXiv:2410.05160},
  year={2024}
}

For additional details regarding the construction process and dataset statistics, please refer to the paper.