MRAMG/MRAMG-Bench

Introduction

MRAMG-Bench is a comprehensive multimodal benchmark with six carefully curated English datasets. The benchmark comprises 4,346 documents, 14,190 images, and 4,800 QA pairs, sourced from three domains—Web Data, Academic Papers, and Lifestyle Data. We believe it provides a robust evaluation framework that advances research in Multimodal Retrieval-Augmented Multimodal Generation (MRAMG).

Data Structure

The dataset consists of three major components: Documents, Multimodal QA pairs, and Images. Each component is structured across six different sub-datasets, ensuring a diverse and comprehensive collection of multimodal content.

---

1. Document Collection

The dataset includes six JSONL files, each corresponding to a different data source:

File Name	Description	Num
doc_wit.jsonl	MRAMG-Wit documents	639
doc_wiki.jsonl	MRAMG-Wiki documents	538
doc_web.jsonl	MRAMG-Web documents	1500
doc_arxiv.jsonl	MRAMG-Arxiv documents	101
doc_recipe.jsonl	MRAMG-Recipe documents	1528
doc_manual.jsonl	MRAMG-Manual documents	40

Field Definitions

• id (int): Unique identifier for the document.
• content (str): The main textual content of the document. If an image is referenced, <PIC> is used as a placeholder indicating its position in the text.
• images_list (list[int]): A list of image IDs associated with the document.

---

2. Multimodal QA pairs

The MQA component consists of six JSONL files, each corresponding to a different dataset:

File Name	Description	Num
wit_mqa.jsonl	MRAMG-Wit multimodal QA pairs	600
wiki_mqa.jsonl	MRAMG-Wiki multimodal QA pairs	500
web_mqa.jsonl	MRAMG-Web multimodal QA pairs	750
arxiv_mqa.jsonl	MRAMG-Arxiv multimodal QA pairs	200
recipe_mqa.jsonl	MRAMG-Recipe multimodal QA pairs	2360
manual_mqa.jsonl	MRAMG-Manual multimodal QA pairs	390

Each entry contains a question ID, a question, provenance documents, a ground truth answer, and a list of image IDs associated with the answer.

Field Definitions

• id (str): Unique identifier for the question.
• question (str): The question text.
• provenance (list[int]): A list of document IDs that serve as supporting evidence for the answer.
• ground_truth (str): The correct answer, which may contain <PIC> placeholders indicating relevant images.
• images_list (list[int]): A list of image IDs directly associated with the answer.

---

3. Image Metadata

The dataset contains a collection of images stored under the directory:

IMAGE/images/

Additionally, metadata about these images is provided in six JSON files, corresponding to each dataset:

File Name	Description	Num
wit_imgs_collection.json	Image metadata from MRAMG-Wit	639
wiki_imgs_collection.json	Image metadata from MRAMG-Web	538
web_imgs_collection.json	Image metadata from MRAMG-Wiki	1500
arxiv_imgs_collection.json	Image metadata from MRAMG-Arxiv	337
recipe_imgs_collection.json	Image metadata from MRAMG-Recipe	8569
manual_imgs_collection.json	Image metadata from MRAMG-Manual	2607

Field Definitions

• id (int): Unique identifier for the image.
• image_url (str): The URL where the image is originally sourced from.
• image_path (str): The filename of the image as stored in the dataset.
• image_caption (str): A textual description or caption of the image.

Contact

If you have any questions or suggestions, please contact [email protected]

Citation Information

If you use this benchmark in your research, please cite the benchmark as follows:

@article{yu2025mramg,
  title={MRAMG-Bench: A BeyondText Benchmark for Multimodal Retrieval-Augmented Multimodal Generation},
  author={Yu, Qinhan and Xiao, Zhiyou and Li, Binghui and Wang, Zhengren and Chen, Chong and Zhang, Wentao},
  journal={arXiv preprint arXiv:2502.04176},
  year={2025}
}