README.md

---
license: mit
task_categories:
- visual-question-answering
- text-to-image
language:
- en
tags:
- vision-language
- vqa
- multimodal
- question-answering
size_categories:
- n<1K
---

# SimpleVQA Dataset

SimpleVQA is a simple vision-language question-answering dataset designed for testing and reproducing vision-language model training. It contains 128 samples with images and question-answer pairs in a conversational format.

## Dataset Description

- **Repository**: [JosephFace/simpleVQA](https://huggingface.co/datasets/JosephFace/simpleVQA)
- **Paper**: N/A
- **Point of Contact**: N/A

### Dataset Summary

SimpleVQA is a lightweight dataset containing 128 vision-language question-answering samples. Each sample includes:
- An image (512x512 RGB)
- A conversation with user questions and assistant answers
- Image paths for reference

This dataset is suitable for:
- Testing vision-language model training pipelines
- Reproducing experimental results
- Educational purposes and quick prototyping

### Supported Tasks

- **Visual Question Answering (VQA)**: Answer questions about image content
- **Image Description**: Generate descriptions of image content
- **Multimodal Conversation**: Engage in conversations about images

### Languages

The dataset is primarily in English.

## Dataset Structure

### Data Fields

Each sample contains the following fields:

- **messages**: List of conversation turns
  - `role`: "user" or "assistant"
  - `content`: Text content of the message
- **image**: PIL Image object (RGB format, 512x512)
- **image_path**: Original image file path

### Data Splits

- **train**: 128 samples

### Example

```python
from datasets import load_dataset

dataset = load_dataset("JosephFace/simpleVQA")

# Access a sample
sample = dataset["train"][0]
print(sample["messages"])
# [
#   {"role": "user", "content": "What is shown in this image?"},
#   {"role": "assistant", "content": "This is sample image 0 from the SimpleVQA dataset."}
# ]

print(sample["image"])  # PIL Image object
print(sample["image_path"])  # "images/image_00000.jpg"
```

## Usage

### Load from HuggingFace Hub

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("JosephFace/simpleVQA")

# Or load specific split
train_dataset = load_dataset("JosephFace/simpleVQA", split="train")
```

### Load Locally

If you have the dataset files locally:

```python
from datasets import load_from_disk

# Load from Arrow format
dataset = load_from_disk("path/to/hf_dataset")

# Or load from JSONL
from datasets import load_dataset
dataset = load_dataset("json", data_files="simpleVQA_128.jsonl", split="train")
```

### Use with Training Pipeline

```python
from datasets import load_dataset
from veomni.data.dataset import MappingDataset

# Load dataset
hf_dataset = load_dataset("JosephFace/simpleVQA", split="train")

# Use with VeOmni training pipeline
dataset = MappingDataset(data=hf_dataset, transform=your_transform_function)
```

## Dataset Statistics

- **Total samples**: 128
- **Image format**: JPEG, 512x512 RGB
- **Average conversation turns**: 2 (1 user question + 1 assistant answer)
- **Total images**: 128

## Limitations

- Small dataset size (128 samples) - suitable for testing only
- Synthetic/placeholder images - not real-world data
- Limited question diversity
- Primarily English language content

## Citation

```bibtex
@dataset{josephface_simplevqa,
  title={SimpleVQA: A Simple Vision-Language Question-Answering Dataset},
  author={JosephFace},
  year={2025},
  url={https://huggingface.co/datasets/JosephFace/simpleVQA}
}
```

## License

This dataset is released under the MIT License.