All notable changes to this project will be documented in this file.
- Bug fixed: Resolved several issues from previous versions.
- Fixed the issue in
ViDoSeek-pageandMMLongBench-page. More details in this issue.
- Release the raw video files here. Please note that raw videos are not required for MMEB evaluation — video frames are all you need. We provide the raw videos only in case they are useful for your own purposes. Please refer to our main data repository for more instructions.
- Updated the MomentSeeker task by deduplicating some test cases. More details in this issue.
- Expanded Modality Support: VLM2Vec now supports unified training and evaluation on three modalities: images, videos, and visual documents.
- VLM2Vec v2: Initial release of the V2 model and framework.
- MMEB-v2 Benchmark: Introduced a new comprehensive benchmark for evaluating performance across all supported modalities.
- Refactored the sub-batch splitting logic within GradCache and the VLM processor to simplify future extensions.