Qhuy204/VQA_VN_Destination

Vietnamese Travel Destination VQA Dataset

Dataset Description

This dataset contains Visual Question Answering (VQA) pairs for Vietnamese travel destinations.

Dataset Statistics

• Total samples: 29759
• Number of shards: 6
• Total size: 9799.20 MB
• Total QA Pairs: ~1,160,600

Files

• train-00000-of-00006.parquet (1486.67 MB, 5000 rows)
• train-00001-of-00006.parquet (1635.00 MB, 5000 rows)
• train-00002-of-00006.parquet (1587.41 MB, 5000 rows)
• train-00003-of-00006.parquet (1784.27 MB, 5000 rows)
• train-00004-of-00006.parquet (1641.47 MB, 5000 rows)
• train-00005-of-00006.parquet (1664.38 MB, 4759 rows)

Data Fields

• id (string): Unique identifier for each sample
• image (Image): Image of Vietnamese travel destination (embedded as bytes)
• conversations (string): JSON string containing question-answer pairs in Vietnamese

Data Format

Each conversations field is a JSON array with role/content format:

[
  {"role": "user", "content": "Nội dung chính của bức ảnh là gì?"},
  {"role": "assistant", "content": "Bức ảnh chụp cảnh đẹp Việt Nam..."}
]

This format is compatible with:

• OpenAI Chat format
• LLaVA training
• LLaMA-style instruction tuning
• Most modern VLM frameworks

Usage

Load with datasets library

from datasets import load_dataset
import json

# Load full dataset
dataset = load_dataset("Qhuy204/VQA_VN_Destination")

# Access and parse conversations
for item in dataset['train']:
    print(item['id'])
    item['image'].show()
    
    # Parse conversations JSON
    conversations = json.loads(item['conversations'])
    for turn in conversations:
        print(f"{turn['role']}: {turn['content']}")

Use with LLaVA/VLM training

from datasets import load_dataset
import json

dataset = load_dataset("Qhuy204/VQA_VN_Destination")

def format_for_llava(example):
    '''Convert to LLaVA training format'''
    convs = json.loads(example['conversations'])
    return {
        'image': example['image'],
        'conversations': convs  # Already in correct format!
    }

# Ready for training
dataset = dataset.map(format_for_llava)

Stream without downloading

from datasets import load_dataset

# Stream mode - memory efficient
dataset = load_dataset("Qhuy204/VQA_VN_Destination", streaming=True)

for item in dataset['train']:
    # Process each item without loading entire dataset
    pass

Load specific shard

from datasets import load_dataset

# Load only first shard
dataset = load_dataset(
    "Qhuy204/VQA_VN_Destination",
    data_files="data/train-00000-of-00006.parquet"
)

Technical Details

• Image Storage: Images are embedded as binary data in Parquet files
• Compression: ZSTD level 3 for optimal size/speed tradeoff
• Sharding: Multiple shards enable parallel downloading and streaming
• Format: Apache Parquet with Hugging Face metadata
• Row Group Size: 100 rows per group for optimal Dataset Viewer performance

Citation

If you use this dataset, please cite:

@dataset{vietnamese_travel_vqa,
  title={Vietnamese Travel Destination VQA Dataset},
  author={Qhuy204},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/Qhuy204/VQA_VN_Destination},
  license={CC BY-SA 4.0},
  source_notes={Images collected from multiple public domain travel blogs and sites, processed by the author.}
}

License

CC BY-SA 4.0

Contact

For questions or issues, please open an issue on the dataset repository.