GridTallyBench: Checkerboard Image Dataset for MLLM Benchmarking

Overview

GridTallyBench is a collection of synthetic checkerboard images designed to test and benchmark Multi-modal Large Language Models (MLLMs) on tasks involving visual pattern recognition and counting. This dataset offers a controlled environment for evaluating model performance on basic visual tasks, particularly useful for assessing an MLLM's ability to count and describe simple geometric patterns.

Dataset Details

• Name: GridTallyBench
• Version: 1.0.0
• Task: Image classification and object counting
• Size: 960 images
• Format: Parquet file containing image data and metadata
• License: MIT

Content

The dataset consists of checkerboard images with the following variations:

• Block sizes: 1x1 to 24x24 pixels
• Grid sizes: 1x1 to 20x20 blocks
• Starting colors: Black-first and white-first patterns

Each image in the dataset is accompanied by metadata including:

• block_pixel: Size of each square in pixels (1 to 24)
• grid_size: Number of squares in each row/column (1 to 20)
• first_block: Color of the top-left square ('black' or 'white')
• image: Binary data of the PNG image

Use Cases

This dataset is particularly useful for:

1. Testing MLLM's ability to count objects in images
2. Evaluating pattern recognition capabilities
3. Assessing color differentiation in simple scenarios
4. Benchmarking performance on controlled, synthetic images

Loading the Dataset

To load and use this dataset with the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("MoonTideF/GridTallyBench")

# Access the first item
first_item = dataset['test'][0]
print(f"Block size: {first_item['block_pixel']}x{first_item['block_pixel']} pixels")
print(f"Grid size: {first_item['grid_size']}x{first_item['grid_size']} blocks")
print(f"First block color: {first_item['first_block']}")
dataset['test'][0]['image'].show()

Dataset Creation

This dataset was generated using a custom Python script. The images are synthetic and do not contain any real-world content or personal information.

Limitations

• The dataset is limited to black and white colors only
• Images are synthetic and may not represent real-world complexity
• The largest image size is 480x480 pixels (20x20 grid with 24x24 pixel blocks)

Citation

If you use this dataset in your research, please cite it as follows:

@misc{gridtallybench,
  author = {MoonTideF},
  title = {GridTallyBench: Checkerboard Image Dataset for MLLM Benchmarking},
  year = {2024},
  publisher = {Hugging Face},
  journal = {Hugging Face Datasets},
  howpublished = {\url{https://huggingface.co/datasets/MoonTideF/GridTallyBench}}
}

Contact

For any questions or feedback regarding this dataset, please contact [Your Contact Information].

---