The dataset that makes MNIST look embarrassingly easy

What it does

Fashion-MNIST contains 60,000 training and 10,000 test images of Zalando articles—28×28 grayscale pixels across ten clothing categories including shirts, sneakers, and ankle boots. It copies the original MNIST dataset’s file format, image size, and train-test split exactly, so existing loaders and pipelines work without modification.

The interesting bit

The maintainers openly admit they want researchers to stop using handwritten digits. Because the format is identical to MNIST, you can swap it into legacy code and watch your 99.7%-accuracy CNN suddenly face a harder problem.

Key highlights

• Drop-in replacement: same 28×28 grayscale dimensions and idx3-ubyte file format as classic MNIST
• Ten labels: T-shirt/top, trouser, pullover, dress, coat, sandal, shirt, sneaker, bag, ankle boot
• Bundled into PyTorch, Keras, TensorFlow, and a dozen other ML libraries natively
• Includes an automated scikit-learn benchmark covering 129 non-deep-learning classifiers
• Community-submitted deep-learning results are collected but explicitly noted as untested by maintainers

Caveats

• The maintainers state they have not verified the community-submitted benchmark results in the leaderboard table.
• Third-party language loaders listed in the README have not all been tested against this dataset.

Verdict

Worth a look if you are still using MNIST as a benchmark and want a slightly harder drop-in substitute. Not for production computer-vision pipelines that need more than 28×28 grayscale thumbnails.