README.md

metadata

license: apache-2.0
task_categories:
  - image-classification
  - unconditional-image-generation
language:
  - en
size_categories:
  - 10K<n<100K

MNIST 28×28 Grayscale Dataset

The original MNIST dataset with handwritten digits in 28×28 grayscale format, stored in efficient Parquet format for modern deep learning applications.

Overview

This dataset contains the original MNIST handwritten digit dataset in its native format:

• Format: 28×28 grayscale images
• Digit labels: 0-9 (single-label classification)
• Image format: Grayscale PIL Images
• Storage: Parquet format for efficient loading

Dataset Statistics

Training Set (60,000 samples)

Digit	Count	Digit	Count
0	5,923	5	5,421
1	6,742	6	5,918
2	5,958	7	6,265
3	6,131	8	5,851
4	5,842	9	5,949

Test Set (10,000 samples)

Digit	Count	Digit	Count
0	980	5	892
1	1,135	6	958
2	1,032	7	1,028
3	1,010	8	974
4	982	9	1,009

Directory Structure

mnist_28/
├── README.md                               # This documentation
├── train-00000-of-00001.parquet            # Training data (Parquet format)
└── test-00000-of-00001.parquet             # Test data (Parquet format)

Key Features

• Original Resolution: Maintains original 28×28 pixel resolution
• Grayscale Format: Single-channel grayscale format as in original MNIST
• PIL Integration: Images loaded as PIL RGB objects ready for preprocessing
• Standard Splits: Maintains original MNIST train/test division
• HuggingFace Compatible: Full integration with datasets library
• Efficient Loading: Parquet format for fast columnar data access and compression

Usage

Loading with HuggingFace Datasets

from datasets import load_dataset

# Load the dataset using the custom script
dataset = load_dataset("FrankCCCCC/mnist_28", trust_remote_code=True)

print(f"Train samples: {len(dataset['train'])}")
print(f"Test samples: {len(dataset['test'])}")

# Access a sample
sample = dataset['train'][0]
print(f"Image shape: {sample['image'].size}")    # (28, 28)
print(f"Image mode: {sample['image'].mode}")     # L (Grayscale)
print(f"Label: {sample['label']}")               # Integer: 0-9

Transformations Applied

The dataset preprocessing pipeline includes:

1. Format Conversion: IDX → Parquet
2. Data Storage: Parquet format for efficient storage and loading
3. Data Type: PIL Image objects for easy integration
4. Format Preservation: Maintains original 28×28 grayscale format

Dataset Format

Each sample contains:

{
    'image': PIL.Image,    # 28×28 grayscale image
    'label': int,          # Digit class (0-9)
}

Technical Details

• Original Source: MNIST Database of Handwritten Digits
• Format: Parquet files for efficient columnar storage
• Preprocessing: Maintains original grayscale format and 28×28 resolution
• Loading: HuggingFace Datasets with custom GeneratorBasedBuilder
• Compression: Parquet format provides built-in compression and fast access

Citation

@article{lecun1998mnist,
  title={The MNIST database of handwritten digits},
  author={LeCun, Yann and Bottou, L{\'e}on and Bengio, Yoshua and Haffner, Patrick},
  year={1998},
  url={http://yann.lecun.com/exdb/mnist/}
}

@misc{mnist28,
  title={MNIST 28×28 Grayscale Dataset},
  author={Original MNIST for Deep Learning},
  year={2024},
  note={Original MNIST dataset in efficient Parquet format}
}

License

This dataset follows the same license as the original MNIST dataset. The original MNIST database is available under the Creative Commons Attribution-Share Alike 3.0 license.

Acknowledgments

• Based on the original MNIST dataset by Yann LeCun, Corinna Cortes, and Christopher J.C. Burges
• Preserved in original format for classical deep learning applications
• Compatible with HuggingFace Datasets ecosystem for seamless integration
• Optimized for CNN architectures and transfer learning applications