README.md

metadata

configs:
  - config_name: docvqa
    default: true
    data_files:
      - split: train
        path: docvqa/train-*.parquet
      - split: test
        path: docvqa/test-*.parquet
  - config_name: papervisa
    data_files:
      - split: train
        path: papervisa/train-*.parquet
      - split: test
        path: papervisa/test-*.parquet
  - config_name: wikivisa
    data_files:
      - split: train
        path: wikivisa/train-*.parquet
      - split: test
        path: wikivisa/test-*.parquet
  - config_name: finewebvisa
    data_files:
      - split: train
        path: finewebvisa/train-*.parquet
  - config_name: visualmrc
    data_files:
      - split: train
        path: visualmrc/train-*.parquet
      - split: val
        path: visualmrc/val-*.parquet
      - split: test
        path: visualmrc/test-*.parquet
  - config_name: visualwebbench
    data_files:
      - split: test
        path: visualwebbench/test-*.parquet
  - config_name: longdocurl
    data_files:
      - split: test
        path: longdocurl/test-*.parquet
  - config_name: mmlongbenchdoc
    data_files:
      - split: test
        path: mmlongbenchdoc/test-*.parquet
  - config_name: doclingmatix
    data_files:
      - split: train
        path: doclingmatix/train-*.parquet

DocAttriBench (DAB)

Welcome to the official Hugging Face page of the DocAttriBench (DAB) dataset, developed in the paper “DocAttriBench: Benchmarking Answer Grounding in Document Visual Question Answering”, accepted at BMVC 2026 (British Machine Vision Conference).

🌐 Project page: https://aimagelab.github.io/DocAttriBench/

This README provides an overview of the dataset and instructions on how to use it.

📊 Dataset Overview

DocAttriBench is a dataset for training and evaluating answer grounding in Document Visual Question Answering. The dataset is constructed from existing document VQA and document understanding datasets. From these source datasets, we use Docling to extract document layout elements and their semantic types, and MAPPET to obtain the bounding box of the evidence region supporting the answer. During dataset construction, we used Qwen2.5-VL-7B-Instruct for answer abstraction and applied MAPPET for source attribution using the perplexity scores produced by this model.

Each item is composed as follows:

Field	Description
image	The document image, stored as a Hugging Face Image feature.
image_path	Original relative image path from the source dataset.
query	The question associated with the document image.
answer	The answer associated with the question, stored in list format.
bbox	The bounding box of the evidence region, stored in the original nested-list format. Each bounding box is represented as normalized coordinates [x1, y1, x2, y2], where (x1, y1) is the top-left corner and (x2, y2) is the bottom-right corner. Values are in [0, 1] and are relative to the image width and height.
type	The semantic type of the evidence region, stored in list format.
source_dataset	Name of the source dataset/configuration.

The evidence-region type can be one of the following:

• paragraph/body
• caption
• heading/title
• subtitle/byline
• data
• sub-data
• image
• picture
• table
• list
• text
• other

⚙️ Dataset Configurations

The dataset is organized using Hugging Face configurations, one for each source dataset. For example:

from datasets import load_dataset

train_docvqa = load_dataset(
    "aimagelab/DocAttriBench",
    "docvqa",
    split="train",
)

Available configurations:

• docvqa
• doclingmatix
• finewebvisa
• longdocurl
• mmlongbenchdoc
• papervisa
• visualmrc
• visualwebbench
• wikivisa

📁 Dataset Splits

The table below reports the number of items available in each split for each configuration.

Configuration	Train	Val	Test
docvqa	4,070	-	467
doclingmatix	139,632	-	-
finewebvisa	33,092	-	-
longdocurl	-	-	688
mmlongbenchdoc	-	-	276
papervisa	52,113	-	1,546
visualmrc	15,000	2,068	4,857
visualwebbench	-	-	233
wikivisa	31,000	-	1,233
Total	274,907	2,068	9,300

In the paper, the validation split is considered part of the training data. In this Hugging Face release, we keep val separate when it is available, so users can decide how to use it.

Please also note that SlideVQA is not included in this release due to licensing constraints. For this reason, the dataset counts in this Hugging Face version differ from the counts reported in the paper.

🧩 Evidence Region Types

The following table reports the number of type annotations for each evidence-region type, keeping train, val, and test separate. Please note that these counts differ from those reported in the paper because SlideVQA is excluded from this Hugging Face release due to licensing constraints.

Type	Train	Val	Test
paragraph/body	146,762	1,480	5,265
caption	15,772	19	347
heading/title	9,881	113	232
subtitle/byline	2,887	73	189
data	1,316	26	68
sub-data	35	10	15
image	810	25	183
picture	6,766	0	433
table	25,649	0	961
list	5,899	112	201
text	4,697	0	891
other	54,451	213	528
Total	274,925	2,071	9,313

⚖️ Dataset Licenses

DocAttriBench is derived from multiple existing datasets, each distributed under its own license. The licenses of the source datasets are:

• DocVQA: Apache License 2.0
• DoclingMatix: Community Data License Agreement – Permissive 2.0
• LongDocURL: Apache License 2.0
• MMLongBench-Doc: Apache License 2.0
• VISA:
  • FineWeb-edu: Open Data Commons Attribution License family
  • NQ: Apache License 2.0
  • PubLayNet: Community Data License Agreement – Permissive, Version 1.0
  • Wikipedia: Creative Commons Attribution-ShareAlike and GNU Free Documentation License family
• VisualMRC: Creative Commons
• VisualWebBench: Apache License 2.0

SlideVQA is not distributed as part of this Hugging Face release, because its license permits usage for testing and evaluation but does not allow redistribution.

Users of DocAttriBench should also comply with the licenses and terms of use of the corresponding source datasets.