Dataset Card for MarineEval

The work introduces MarineEval, the first large-scale benchmark specifically designed to evaluate the marine understanding capabilities of Vision-Language Models (VLMs). MarineEval contains 2,000 expert-verified image-based question–answer pairs across 7 task dimensions and 20 domain-specific capacity dimensions, emphasizing specialized marine knowledge, visual reasoning, and real-world complexity. Through comprehensive benchmarking of 17 state-of-the-art VLMs, the study reveals that existing general-purpose models perform poorly on marine tasks, including particularly in spatial reasoning, species identification, and ecological understanding, highlighting the need for domain-aware training and evaluation. This resource aims to foster progress toward domain-expert VLMs capable of advancing research and conservation in marine science.

Dataset

Dataset Structure

The dataset structure is as follows:

dataset/
├── dimension 1
│   ├── sub dimension 1
│   │   ├── images/
│   │   ├── data.json
│   ├── sub dimension 1
│   │   ├── images/
│   │   ├── data.json
├── dimension 2
│   ├── sub dimension 1
│   │   ├── images/
│   │   ├── data.json
...

JSON File Structure

Each data.json file follows this structure:

"data": [
    {
        "id": 0,
        "question": "string",
        "answers": [
            {
                "answer": "string",
            }
        ],
        "qusetion_format": 0
    }
]

Question Formats

The MarineEval dataset includes five types of question formats:

Code	Question Format	Description
0	Yes-No Question	Models make binary classification to determine whether a statement is true or false.
1	Multiple Choice Question	Models select one or more than one correct option from at least four choices
2	Summarization Question	Models are asked to summarize the insight of the given image in free format
3	Localization Question	Models are asked to provide bounding box of target objects in COCO format.
4	Closed-Form (Loose)	Models response in a restricted format, which is evaluated with flexible semantic matching by LLM.
5	Closed-Form (Strict)	Models respond in a restricted format, which requires an exact match with the ground truth.

Citation

@misc{wong2025marineevalassessingmarineintelligence,
      title={MarineEval: Assessing the Marine Intelligence of Vision-Language Models}, 
      author={YuK-Kwan Wong and Tuan-An To and Jipeng Zhang and Ziqiang Zheng and Sai-Kit Yeung},
      year={2025},
      eprint={2512.21126},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.21126}, 
}