README.md

metadata

language:
  - vi
  - en
task_categories:
  - visual-question-answering
  - question-answering
tags:
  - infographic
  - vietnamese
  - vqa
  - document-understanding
size_categories:
  - 10K<n<100K

ViInfographicVQA

Overview

ViInfographicVQA is a Vietnamese Visual Question Answering (VQA) benchmark for infographic understanding.
It evaluates models’ ability to read, reason, and synthesize information from data-rich, layout-heavy visuals that mix text, charts, maps, and design elements.

Two settings are provided:

• Single-image VQA – questions answered from one infographic.
• Multi-image VQA – questions requiring reasoning across multiple, semantically related infographics.

---

📊 Dataset Summary

Split	#Images	#QAs	Description
Single-image (train)	1,787	12,521	VQA on individual infographics
Single-image (test)	193	1,374	Held-out evaluation
Multi-image (train)	5,861	5,878	Cross-image reasoning (training)
Multi-image (test)	653	636	Cross-image reasoning (test)
Total	6,747	20,409	Across all splits

• Language: Vietnamese
• Domains: Economy, Healthcare, Education, Society & Culture, Disasters & Accidents, Sports & Arts, Weather, etc.

🗂️ Repository Layout

ViInfographicVQA/
├── images/                # all image files (referenced by filename)
├── <parquet files>        # four splits stored as parquet shards on the Hub
└── README.md

🚀 Quickstart

from datasets import load_dataset

# Load all splits (parquet)
ds = load_dataset("VLAI-AIVN/ViInfographicVQA")

single_train = ds["single_train"]
multi_train  = ds["multi_train"]

# Each sample:
# - images_paths: list of filenames (relative to `images/`)
# - image: preview Image() (the first file)
ex = multi_train[0]
print(ex["images_paths"])  # e.g. ["13321.jpg", "13028.jpg", "13458.jpg"]
preview = ex["image"]      # PIL.Image preview (for quick visualization)

Read all images for multi-image samples (no local download)

Use Hub file URIs, then cast to Image():

from datasets import Image, Sequence, load_dataset

ds = load_dataset("VLAI-AIVN/ViInfographicVQA")
repo_base = "hf://datasets/VLAI-AIVN/ViInfographicVQA/images"

def add_full_paths(example):
    example["images_full"] = [f"{repo_base}/{fn}" for fn in example["images_paths"]]
    return example

multi = ds["multi_train"].map(add_full_paths, remove_columns=[])
multi = multi.cast_column("images_full", Sequence(Image()))

all_imgs = multi[0]["images_full"]   # list[PIL.Image] — all referenced images

Streaming (large-scale training)

from datasets import load_dataset, Image, Sequence

ds = load_dataset("VLAI-AIVN/ViInfographicVQA", streaming=True)
repo_base = "hf://datasets/VLAI-AIVN/ViInfographicVQA/images"

def add_full_paths(example):
    example["images_full"] = [f"{repo_base}/{fn}" for fn in example["images_paths"]]
    return example

multi_stream = ds["multi_train"].map(add_full_paths)
multi_stream = multi_stream.cast_column("images_full", Sequence(Image()))

ex = next(iter(multi_stream))
imgs = ex["images_full"]  # list of PIL.Image (lazy/streamed)

Local download (offline use)

from huggingface_hub import snapshot_download
from datasets import load_dataset

# Download the entire dataset repo locally (parquet + images)
local_dir = snapshot_download(repo_id="VLAI-AIVN/ViInfographicVQA", repo_type="dataset")

# Load from disk
ds = load_dataset(local_dir)

# Reconstruct absolute paths to images on disk if needed:
import os
images_root = os.path.join(local_dir, "images")
def to_abs(example):
    example["images_abs"] = [os.path.join(images_root, fn) for fn in example["images_paths"]]
    return example

multi_local = ds["multi_train"].map(to_abs)
print(multi_local[0]["images_abs"][:3])  # ['/.../images/13321.jpg', ...]

Speed tip: set HF_HUB_ENABLE_HF_TRANSFER=1 to accelerate uploads/downloads.

🔍 Research Applications

• Multimodal reasoning on charts, tables, and dense text
• Cross-image synthesis and comparison
• Low-resource VQA in Vietnamese
• Evaluation of OCR, layout parsing, and numerical reasoning

🧮 Evaluation

We use Average Normalized Levenshtein Similarity (ANLS) for string-based answer evaluation, which tolerates minor textual variations while penalizing semantic errors.

📚 Citation

If you use this dataset, please cite:

@article{van2025viinfographicvqa,
  title={ViInfographicVQA: A Benchmark for Single and Multi-image Visual Question Answering on Vietnamese Infographics},
  author={Van-Dinh, Tue-Thu and Tran, Hoang-Duy and Duong, Truong-Binh and Pham, Mai-Hanh and Le-Nguyen, Binh-Nam and Nguyen, Quoc-Thai},
  journal={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}