VietTravelVQA

VietTravelVQA is a Vietnamese visual question answering dataset focused on tourism and cultural heritage in Vietnam. It contains 1,406 images and 7,030 human-verified question-answer pairs organized into five difficulty levels, from basic visual recognition to cultural reasoning.

The authoritative release is archived on Mendeley Data: VietTravelVQA, version 2.

Dataset summary

Split	Conversations	Question-answer pairs	Images
Train	6,094	6,094	1,406 unique images
Test	936	936	678 unique images
Total	7,030	7,030	1,406 unique images

The split is defined at the question-answer level. An image may therefore be referenced by examples in both splits.

Files and formats

data/train-00000-of-00049.parquet  # Train shards (49 files)
data/test-00000-of-00008.parquet   # Test shards (8 files)

Annotation format

The Parquet files contain four columns: messages, images, difficult, and source. Each row contains exactly one image and one user-assistant question-answer pair. The image bytes are embedded directly in the images column, and difficult stores the difficulty level for that pair.

{
  "messages": [
    {
      "role": "user",
      "content": [
        {"index": null, "text": "Có bao nhiêu trụ cầu có thể nhìn thấy rõ trong hình?", "type": "text"},
        {"index": 0, "text": null, "type": "image"}
      ]
    },
    {
      "role": "assistant",
      "content": [
        {"index": null, "text": "Bốn.", "type": "text"}
      ]
    }
  ],
  "images": ["<embedded image>"],
  "difficult": "2",
  "source": {
    "platform": "Wikimedia Commons",
    "license": "CC-BY-SA (Wikimedia Commons)",
    "url": "https://upload.wikimedia.org/..."
  }
}

Difficulty levels

The source release covers five levels of complexity, ranging from direct perception and object identification to spatial, historical, and culturally grounded reasoning.

Data collection and validation

Images were sourced from Wikimedia Commons. Expert annotators created Vietnamese question-answer pairs concerning architectural details, historical significance, and spatial reasoning. Automated consistency checks synchronized the annotations with image files, and Large Multimodal Models were used to assist factual auditing, followed by human verification.

Intended uses

The dataset is intended for training, fine-tuning, and evaluating vision-language models on Vietnamese tourism and cultural heritage content. Potential uses include visual question answering, localized cultural reasoning, and evaluation across difficulty levels.

Limitations

• Coverage is centered on tourism and cultural heritage in Vietnam and should not be treated as representative of all Vietnamese visual-language use cases.
• Model-assisted auditing does not guarantee that every annotation is free of factual or linguistic errors.
• The question-answer split may reference the same image in both train and test. Evaluation protocols requiring image-disjoint splits should create a new split grouped by the embedded image.

License and attribution

The dataset release is licensed under Creative Commons Attribution 4.0 International. Individual images originate from Wikimedia Commons and retain their source license and attribution requirements. Consult the authoritative Mendeley Data release for the original per-image source metadata before redistributing or reusing an image.

Citation

If you use VietTravelVQA, please cite the Mendeley Data release:

@dataset{viettravelvqa_2026,
  author    = {Nguyen Van Nha and Phung The Huan and Phu Nguyen Dinh and
               Le Hong Quan and Truong Quoc Huy and Pham Khanh Duy and
               Le Minh Tuan and Le Hoang Son},
  title     = {VietTravelVQA},
  year      = {2026},
  publisher = {Mendeley Data},
  version   = {2},
  doi       = {10.17632/fvxtpnh8mh.2},
  url       = {https://doi.org/10.17632/fvxtpnh8mh.2}
}

Contributors

• Nguyen Van Nha
• Phung The Huan
• Phu Nguyen Dinh
• Le Hong Quan
• Truong Quoc Huy
• Pham Khanh Duy
• Le Minh Tuan
• Le Hoang Son