ZHNie/MBE2.0

This is the HuggingFace repository of the paper named MOON2.0: Dynamic Modality-balanced Multimodal Representation Learning for E-commerce Product Understanding.

Following internal legal, privacy, and compliance review (aligned with China’s PIPL), we have released the MBE2.0 benchmark, including original images, titles and category/attribute annotations. All personally identifiable information has been rigorously removed. The training set and model weights are undergoing final security clearance, with a commitment to full public release.

The MBE2.0 Benchmark is a co-augmented multimodal representation benchmark designed specifically for representation learning and evaluation in e-commerce scenarios.

Our data originates from Taobao, one of the largest e-commerce platforms in China. By collecting and processing user interaction logs spanning from January 1, 2023, to June 30, 2025, we have constructed dedicated training and test sets.

Following internal open-source approval procedures, we hereby release the complete test set of the MBE2.0 Benchmark. Due to repository size limitations imposed by the anonymous submission system, we provide a representative subset of 10,000 samples. The full training set and associated model checkpoints are currently undergoing internal review. We commit to releasing the complete dataset and models.

This benchmark is for academic research use only and is prohibited from use in any commercial setting.

@article{nie2025moon2,
  title={MOON2. 0: Dynamic Modality-balanced Multimodal Representation Learning for E-commerce Product Understanding},
  author={Nie, Zhanheng and Fu, Chenghan and Zhang, Daoze and Wu, Junxian and Guan, Wanxian and Wang, Pengjie and Xu, Jian and Zheng, Bo},
  journal={arXiv preprint arXiv:2511.12449},
  year={2025}
}