Y-J-Ju/MMEB-train

Massive Multimodal Embedding Benchmark

The training data split used for training VLM2Vec models in the paper VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks (ICLR 2025).

MMEB benchmark covers 4 meta tasks and 36 datasets meticulously selected for evaluating capabilities of multimodal embedding models.

During training, we utilize 20 out of the 36 datasets.

For evaluation, we assess performance on the 20 in-domain (IND) datasets and the remaining 16 out-of-domain (OOD) datasets. Please refer to TIGER-Lab/MMEB-eval for the test split of MMEB.

News

[2025-01]: We have updated our training data. Each subset now contains two splits: original and diverse_instruction. The original split is provided to support the reproduction of our paper results. The diverse_instruction split includes paraphrased instructions for each task, designed to enhance instruction diversity and improve the model's robustness to unseen instructions and tasks. Moving forward, our future releases will primarily use the diverse_instruction split.

Dataset Usage

For each dataset, we have 1000 examples for evaluation. Each example contains a query and a set of targets. Both the query and target could be any combination of image and text. The first one in the candidate list is the groundtruth target.

Statistics

We show the statistics of all the datasets as follows:

Cite Us

@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}
}