Datasets:

moca-embed
/

MMEB-train-fixed

dataset_info:
  features:
    - name: text
      dtype: string
    - name: images
      sequence: binary
  splits:
    - name: train
      num_bytes: 94020556918
      num_examples: 1465964
  download_size: 73033984223
  dataset_size: 94020556918
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

MMEB train split used in MoCa Continual Pre-training

Introduction

This is a interleaved multimodal pre-training dataset used in the modality-aware continual pre-training of MoCa models. It is adapted from the train split of MMEB by concatenating queries and positive documents.

The dataset consists of interleaved multimodal examples. text is a string containing text while images are image binaries that can be loaded with the following code snippet:

import PIL.Image
from io import BytesIO

image_bytes = example['images'][0]
image = PIL.Image.open(BytesIO(image_bytes))

Citation

MoCa

@article{chen2025moca,
  title={MoCa: Modality-aware Continual Pre-training Makes Better Bidirectional Multimodal Embeddings},
  author={Chen, Haonan and Liu, Hong and Luo, Yuping and Wang, Liang and Yang, Nan and Wei, Furu and Dou, Zhicheng},
  journal={arXiv preprint arXiv:2506.23115},
  year={2025}
}

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}
}