Kuzushiji-MNIST (KMNIST) is a dataset of 70,000 grayscale images of handwritten Japanese Hiragana characters, each 28x28 pixels. It was created as a more challenging alternative to the classic MNIST dataset, which contains handwritten digits. KMNIST is designed to be a drop-in replacement for MNIST in machine learning benchmarks, allowing researchers to test models on a task with similar structure but greater visual complexity due to the cursive nature of Hiragana.
The dataset was introduced in 2018 by a team of researchers at the University of Tokyo, including Tarin Clanuwat, Mikel Bober-Irizar, Asanobu Kitamoto, and Alex Lamb. It was derived from the Kuzushiji dataset, which contains over 140,000 images of classical Japanese literature. The KMNIST subset focuses on 10 Hiragana characters, each represented by 7,000 images (6,000 for training and 1,000 for testing), mirroring the exact split of MNIST.
Dataset Structure
KMNIST follows the same format as MNIST: each image is a 28x28 pixel grayscale array, with pixel values ranging from 0 to 255. The dataset is split into 60,000 training images and 10,000 test images. The 10 characters included are: 'o', 'ki', 'su', 'tsu', 'na', 'ha', 'ma', 'ya', 're', and 'wo'. These characters were selected because they are among the most common in the original Kuzushiji dataset and present a balanced classification challenge.
The creators intentionally avoided using the same characters as MNIST digits, ensuring that models trained on KMNIST cannot simply transfer learned features from digit recognition. The Hiragana characters have more complex stroke patterns and curvatures, making the task harder for traditional machine learning models.
Benchmarking and Performance
KMNIST has become a standard benchmark in the Machine learning community, particularly for evaluating Neural network architectures. On the original MNIST, many models achieve over 99% accuracy, but KMNIST is significantly harder. For example, a simple Residual Network (ResNet) might achieve around 99% on MNIST but only around 95% on KMNIST. This makes KMNIST useful for distinguishing between models that perform well on simple tasks and those that generalize to more complex patterns.
Researchers have used KMNIST to test various techniques, including Data Augmentation, Dropout, and Batch Normalization. The dataset is also included in popular machine learning libraries such as PyTorch and TensorFlow, making it easy to integrate into existing pipelines.
Related Datasets
KMNIST is part of a family of MNIST-like datasets, including Fashion-MNIST (clothing items) and EMNIST (letters and digits). These datasets share the same image format and split, allowing for direct comparison of model performance across different domains. KMNIST is unique in that it focuses on a non-Latin script, providing a cultural and linguistic challenge that other datasets do not.
The original Kuzushiji dataset, from which KMNIST is derived, includes over 4,000 different characters and is used for historical document analysis. KMNIST simplifies this to a 10-class problem, making it accessible for educational purposes and quick experimentation.
Applications and Impact
KMNIST has been widely adopted in Deep learning research and education. It is often used in introductory courses to teach Convolutional neural network design, as it requires more sophisticated architectures than MNIST. The dataset has also been used in studies on Transfer learning and domain-adaptation, where models pretrained on MNIST are fine-tuned on KMNIST.
Beyond benchmarking, KMNIST has contributed to the preservation and study of Japanese cultural heritage. By making classical Hiragana characters accessible in a machine-readable format, it enables automated analysis of historical documents, which is valuable for historians and linguists.
Limitations and Future Directions
Despite its utility, KMNIST has limitations. The 10 characters represent only a small fraction of the full Hiragana syllabary, and the images are preprocessed to match MNIST's format, which may lose some of the original stroke information. Additionally, the dataset is relatively small compared to modern large-scale datasets, which can limit its use for training very deep models.
Future work may involve expanding KMNIST to include more characters or creating variants with different preprocessing. As of 2025, KMNIST remains a standard benchmark, but researchers are increasingly turning to more complex datasets like CIFAR-10 or ImageNet for evaluating state-of-the-art models. Nevertheless, KMNIST continues to serve as a valuable tool for quick prototyping and educational purposes.