Haon-Chen/mmE5-MMEB-hardneg

mmE5 Labeled Data

This dataset contains datasets used for the supervised finetuning of mmE5 (mmE5: Improving Multimodal Multilingual Embeddings via High-quality Synthetic Data):

• MMEB (with hard negative)
• InfoSeek (from M-BEIR)
• TAT-DQA
• ArxivQA

Github

Image Preparation

First, you should prepare the images used for training:

Image Downloads

• Download Links: Download image resources for each dataset via the following links:
  • MMEB
  • InfoSeek
  • ArxivQA
  • TAT-DQA

For TAT-DQA, you need to first save images into the overall image folder to align usage:

dataset = load_dataset(
    "vidore/tatdqa_train",
    split="train"
)
image_out_dir = "images/TAT-DQA"
os.makedirs(image_out_dir, exist_ok=True)
for i, sample in enumerate(dataset):
  save_path = os.path.join(image_out_dir, f"tatdqa_{i}.png")
  if os.path.exists(save_path):
    continue
  image = sample["image"]
  image.save(save_path, format="PNG")

Image Organization

  images/
  ├── mbeir_images/
  │     └── oven_images/
  │           └── ... .jpg (InfoSeek)
  ├── ArxivQA/
  │     └── images/
  │           └── ... .jpg (ArxivQA)
  └── TAT-DQA/
  │     └── ... .png (TAT-DQA)
  └── A-OKVQA/
        └── Train/
  │           └── ... .jpg (A-OKVQA)
  │
  ... (MMEB Training images)

You can refer to the image paths in each subset to view the image organization.

You can also customize your image paths by altering the image_path fields.

Citation

If you use this dataset in your research, please cite the associated paper.

@article{chen2025mmE5,
  title={mmE5: Improving Multimodal Multilingual Embeddings via High-quality Synthetic Data},
  author={Chen, Haonan and Wang, Liang and Yang, Nan and Zhu, Yutao and Zhao, Ziliang and Wei, Furu and Dou, Zhicheng},
  journal={arXiv preprint arXiv:2502.08468},
  year={2025}
}