Datasets:
Multi-MMBench
Multi-MMBench is the multilingual version of MMBench, a multiple-choice benchmark for evaluating vision-language models (VLMs). It is released in English, Italian, French, German and Spanish, and contains 21,645 entries (4,329 per language).
The dataset is part of MEVBench (Multimodal European Vision Benchmark) and was used as an evaluation benchmark in the paper Multilingual Training and Evaluation Resources for Vision-Language Models.
Source dataset
| Dataset | MMBench |
| Hugging Face | lmms-lab-encoder/MMBench |
| Paper | MMBench: Is Your Multi-modal Model an All-around Player? (Liu et al., 2023) |
| Official repository | open-compass/MMBench |
MMBench is a bilingual (English/Chinese) multiple-choice benchmark that evaluates vision-language models across 20 fine-grained ability dimensions (six ability groups), using a CircularEval strategy that rotates the answer options to remove position bias. The Hugging Face release used here (lmms-lab-encoder/MMBench, configuration en, split dev) contains 4,329 rows: the 1,164 MMBench-Dev questions expanded into the option permutations required by CircularEval.
How the dataset was built
The English dev split was extended to four additional languages through an automated translation pipeline, keeping the benchmark structure intact:
- Translated fields:
question, the answer optionsA–D, andhint. - Translation model:
gemini-3-flash-preview. - Unchanged: images, row order, the bookkeeping fields (
index,comment,source,split), the gold answer (answer, i.e. the option letter) and the ability labels (category,l2-category). Since the option letters and their order are preserved, the gold answers remain valid and results are directly comparable across languages. - English configuration: the
enconfiguration contains the original English rows and is kept as a reference for cross-lingual comparison. - No regeneration or re-annotation: only the textual content was translated; images were not modified or regenerated.
The translation protocol is described in Section 4 of the accompanying paper.
Dataset structure
Configurations and splits
Each language is a separate configuration, and all configurations share the same split name (validation, corresponding to the dev split of the source release).
| Config | Language | Split | Entries |
|---|---|---|---|
en |
English | validation | 4,329 |
fr |
French | validation | 4,329 |
de |
German | validation | 4,329 |
it |
Italian | validation | 4,329 |
es |
Spanish | validation | 4,329 |
| Total | 21,645 |
Fields
| Field | Description |
|---|---|
index |
Row index inherited from the source release |
question |
Question text (translated) |
hint |
Optional hint or additional context for the question (translated) |
A, B, C, D |
Answer options (translated; letters unchanged) |
answer |
Gold answer: the letter of the correct option |
category |
One of the 20 MMBench ability dimensions (unchanged, in English) |
l2-category |
One of the six MMBench ability groups (unchanged, in English) |
image |
Image associated with the question |
source |
Provenance of the image in the source benchmark |
comment |
Bookkeeping field inherited from the source release (e.g. source image file name) |
split |
Original MMBench split label |
Usage
from datasets import load_dataset
ds = load_dataset("VillanovaAI/Multi-MMBench", "it", split="validation")
print(ds[0]["question"])
print(ds[0]["A"], ds[0]["B"], ds[0]["C"], ds[0]["D"])
print(ds[0]["answer"])
Limitations
- The translations are produced automatically and inherit any error of the source benchmark as well as any error of the translation model. They are intended for evaluation, not for training.
- Ability labels, provenance fields and identifiers are kept in English on purpose, to preserve comparability with the original benchmark.
Licensing and attribution
This multilingual release is distributed under ODC-BY, as declared in the dataset repository metadata. The upstream benchmark is MMBench (OpenCompass community). Images are redistributed unchanged, while only the textual fields are translated; the conditions of the original image providers continue to apply.
Citation
If you use this dataset, please cite the paper describing the multilingual benchmark suite and the original MMBench paper.
@article{baiamonte2026multilingual,
title={Multilingual Training and Evaluation Resources for Vision-Language Models},
author={Baiamonte, Daniela and Fano, Elena and Gabburo, Matteo and Simonazzi, Stefano and Rigutini, Leonardo and Zugarini, Andrea},
journal={arXiv preprint arXiv:2604.18347},
year={2026}
}
@article{liu2023mmbench,
title={MMBench: Is Your Multi-modal Model an All-around Player?},
author={Liu, Yuan and Duan, Haodong and Zhang, Yuanhan and Li, Bo and Zhang, Songyang and Zhao, Wangbo and Yuan, Yike and Wang, Jiaqi and He, Conghui and Liu, Ziwei and Chen, Kai and Lin, Dahua},
journal={arXiv preprint arXiv:2307.06281},
year={2023}
}
- Downloads last month
- 88