Generative-VQA-V2-Curated

A curated, balanced, and cleaned version of the VQA v2 dataset specifically optimized for Generative Visual Question Answering.

This dataset transforms the standard VQA task into a generative challenge by removing "yes/no" shortcuts and balancing answer distributions to prevent model over-fitting on dominant classes.

Dataset Summary

The primary goal of this curated set is to provide a "clean" signal for training multimodal models by:

• Eliminating Binary Biases: Removed all "yes/no" and "unknown" style answers
• Balancing Classes: Capped samples at 600 per answer to ensure the model learns a diverse vocabulary
• Filtering Ambiguity: Removed generic questions (e.g., "What is this?") to focus on specific visual grounding

Dataset Statistics

• Total QA Pairs: 135,268
• Unique Answer Classes: 1,251
• Source Images: COCO Train 2014
• Minimum Frequency per Answer: 20
• Maximum Samples per Answer: 600
• Average Question Length: ~6 words
• Average Answer Length: ~1.5 words

Curation Logic

The dataset was generated using the following filtering pipeline:

1. Consensus-Based: Only the majority-vote answer from the 10 human annotators is used
2. Exclusion List:
   • Boolean answers: yes, no
   • Uncertainty markers: unknown, none, n/a, cant tell, not sure
3. Ambiguity Filter: Removed questions containing:
   • "what is in the image"
   • "what is this"
   • "what is that"
   • "what do you see"
4. Conciseness: Answers are restricted to ≤5 words and ≤30 characters

Repository Structure

Deva8/Generative-VQA-V2-Curated/
├── main_metadata.csv         # ⭐ Primary data file (17 MB)
├── gen_vqa_v2-images.zip     # 📦 Images archive (10.1 GB)
└── README.md

Inside gen_vqa_v2-images.zip:

gen_vqa_v2-images.zip (10.1 GB)
└── gen_vqa_v2-images/
    └── gen_vqa_v22/
        └── images/
            ├── COCO_train2014_000000004702.jpg
            ├── COCO_train2014_000000012460.jpg
            ├── COCO_train2014_000000183672.jpg
            └── ... (135,268 images total)

Note: The zip also contains metadata.csv and qa_pairs.json files which are not used by this dataset. Please use main_metadata.csv at the repository root instead.

Download Instructions

Option 1: Using huggingface_hub (Recommended)

from huggingface_hub import hf_hub_download
import zipfile
import os

# Download the images zip file (10.1 GB - will be cached)
zip_path = hf_hub_download(
    repo_id="Deva8/Generative-VQA-V2-Curated",
    filename="gen_vqa_v2-images.zip",
    repo_type="dataset"
)

# Extract to a directory
extract_dir = "./gen_vqa_images"
os.makedirs(extract_dir, exist_ok=True)

print(f"Extracting {zip_path}...")
with zipfile.ZipFile(zip_path, 'r') as zip_ref:
    zip_ref.extractall(extract_dir)

print(f"✓ Images extracted to: {extract_dir}")
print(f"✓ Found {len([f for f in os.listdir(os.path.join(extract_dir, 'gen_vqa_v2-images/gen_vqa_v22/images')) if f.endswith('.jpg')])} images")

Option 2: Manual Download

1. Go to: https://huggingface.co/datasets/Deva8/Generative-VQA-V2-Curated/tree/main
2. Click on gen_vqa_v2-images.zip (10.1 GB)
3. Click the download button
4. Extract the zip file to your working directory

🔧 Metadata Fields

The dataset viewer above shows the metadata CSV with the following columns:

Field	Type	Description
image_id	int64	Original COCO Image ID
question_id	int64	Original VQA v2 Question ID
question	string	Natural language question about the image
answer	string	Curated ground-truth answer
file_name	string	Relative path to image file

Example Rows:

image_id,question_id,question,answer,file_name
429568,429568000,What is behind the street sign?,tree,gen_vqa_v2-images/gen_vqa_v22/images/COCO_train2014_000000429568.jpg
4702,4702000,What is on the man's head?,soccer ball,gen_vqa_v2-images/gen_vqa_v22/images/COCO_train2014_000000004702.jpg
183672,183672001,How old is the man?,20,gen_vqa_v2-images/gen_vqa_v22/images/COCO_train2014_000000183672.jpg

📜 License & Attribution

This dataset is a derivative work of:

• VQA v2 Dataset (Goyal et al., 2017) - CC BY 4.0
• COCO Dataset (Lin et al., 2014) - CC BY 4.0

All derivative work is released under the same MIT License.

Original Papers:

@inproceedings{goyal2017making,
  title={Making the v in vqa matter: Elevating the role of image understanding in visual question answering},
  author={Goyal, Yash and Khot, Tejas and Summers-Stay, Douglas and Batra, Dhruv and Parikh, Devi},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={6904--6913},
  year={2017}
}

@inproceedings{lin2014microsoft,
  title={Microsoft coco: Common objects in context},
  author={Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Doll{\'a}r, Piotr and Zitnick, C Lawrence},
  booktitle={European conference on computer vision},
  pages={740--755},
  year={2014},
  organization={Springer}
}

📖 Citation

If you use this dataset in your research or project, please cite:

@misc{devarajan_genvqa_2026,
  author = {Devarajan},
  title = {Generative-VQA-V2-Curated: A Balanced Dataset for Open-Ended Generative VQA},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Deva8/Generative-VQA-V2-Curated}}
}

🤝 Contributing

Found an issue or have suggestions? Please open a discussion on the HuggingFace dataset page!