---
pretty_name: "BToks MMEB-V2"
language:
- en
tags:
- bottleneck-tokens
- btoks
- multimodal-retrieval
- lance
- reproducibility
task_categories:
- image-to-text
- visual-question-answering
- text-retrieval
license: other
viewer: false
---

# BToks MMEB-V2

This dataset repository contains Lance-format converted data used by the open-source reproduction code for **Bottleneck Tokens for Unified Multimodal Retrieval** (arXiv:2604.11095).

## Source

Converted from MMEB-V2 evaluation sources, including [ziyjiang/MMEB_Test_Instruct](https://huggingface.co/datasets/ziyjiang/MMEB_Test_Instruct) and the public upstream datasets referenced by MMEB-V2.. This repository does not change upstream ownership, licensing, citation requirements, or usage restrictions.

## Format

The data is stored as Lance tables for the BToks/VLM2Emb training and evaluation scripts. The Hugging Face Dataset Viewer is disabled because this repository is not stored as standard Parquet or JSON splits.

## Usage

Use this repository ID in the corresponding dataset config:

```yaml
dataset_name: siyrus/BToks-MMEB-V2
```

See the public BToks code repository for the config schema, training presets, and evaluation commands: https://github.com/siryuson/BottleneckTokens

## License and Citation

Users must follow the license, citation requirements, and usage restrictions of the upstream source dataset.

```bibtex
@article{sun2026bottlenecktokens,
  title={Bottleneck Tokens for Unified Multimodal Retrieval},
  author={Sun, Siyu and others},
  journal={arXiv preprint arXiv:2604.11095},
  year={2026}
}
```