# MNIST Dataset Release

The MNIST dataset is a widely used benchmark of 70,000 handwritten digit images, created by remixing NIST datasets, released in 1998 for training machine learning models in image recognition.

The MNIST database (Modified National Institute of Standards and Technology database) is a large collection of handwritten digit images, commonly used for training and testing image processing and machine learning systems. Created by re-mixing samples from original NIST datasets, it was designed to address the limitations of NIST's training data, which came from American Census Bureau employees, and its testing data, from American high school students, making it unsuitable for machine learning experiments. The black-and-white images were normalized to 28x28 pixel bounding boxes and anti-aliased, introducing grayscale levels.

MNIST contains 60,000 training images and 10,000 testing images, with half of each set drawn from NIST's training data and half from its testing data. The creators maintain a list of methods tested on the dataset, and their original paper reported a 0.8% error rate using a support-vector machine. The original MNIST dataset includes at least four incorrect labels, a well-known quirk.

## Predecessors and Origins

Before MNIST, the USPS database (1988) provided a smaller digit dataset from handwritten zip codes on U.S. mail. It contained 16x16 grayscale images, with 7,291 training and 2,007 test images, totaling 9,298. This dataset included ambiguous and misclassified samples, and was challenging, with human error rates averaging 2.5%. It was used to train the early 1989 LeNet architecture.

In the late 1980s, the Census Bureau sought to automate handwritten form digitization, hiring NIST's Image Recognition Group to evaluate OCR systems. Their work produced several Special Databases: SD-1 (1990), SD-3 (1992), and SD-7 (1992). These were built from Handwriting Sample Forms (HSFs), scanned at 300 dpi, with various fields for names, digits, and letters. SD-3 contained 128x128 binary images of segmented alphanumericals, including 223,125 digits, and was derived from census workers. SD-7, the test set, had 58,646 images from 500 high school students, with a human error rate of 1.5%. Notably, SD-3 was cleaner and easier, but machine learning systems trained on SD-3 saw performance drops to ~-10% error on SD-7, highlighting distribution shifts.

A 1992 NIST competition used SD-3 as training data and SD-7 as the test set, with 45 algorithms from 26 companies. The winning entry used a proprietary training set, while the best among SD-3-trained systems was a nearest-neighbor classifier with an invariant metric. SD-19 (1995) combined several special databases, containing 814,255 images.

## Creation of MNIST

Sometime before summer 1994, MNIST was constructed from SD-3 and SD-7. The creators divided SD-7's images into training and test sets, each from 250 writers, yielding nearly 30,000 images. They then augmented each with SD-3 images to reach 60,000 samples per set. All images were size-normalized to the 20x20 pixel box, preserving aspect ratio, and anti-aliased to grayscale, then centered in a 28x28 image by adjusting to center-of-mass. The downsampling details were later reconstructed.

Initially, the training and test sets both had 60,000 samples, but to reduce test set size, 50,000 samples were discarded, keeping only indices 24476 to 34475, resulting in 10,000 test images.

## Evaluation and Impact

MNIST became a standard baseline in [machine-learning](https://www.wikiprompt.org/wiki/machine-learning) and [deep-learning](https://www.wikiprompt.org/wiki/deep-learning) research. It is widely used to compare algorithms, from simple linear models to complex [neural-network](https://www.wikiprompt.org/wiki/neural-network) models. The dataset's small image size and modest number of classes make it ideal for prototyping. Its public availability and clear benchmarks have made it a common educational tool.

## Legacy and Successors

Extended MNIST (EMNIST) is a newer dataset developed by NIST, released in 2017, as a successor to MNIST. It includes handwritten digits and letters, with broader categories. In 2019, the full 60,000 test set of MNIST was restored to create QMNIST, providing a larger test set for research. These derivatives keep MNIST's structure while expanding its diversity.

## See Also

* [artificial-intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence)
* [computer-vision](https://www.wikiprompt.org/wiki/computer-vision)
* [optical-character-recognition](https://www.wikiprompt.org/wiki/optical-character-recognition)

## References

* Original MNIST paper by Yann LeCun et al., 1998 (provided in source facts).
* NIST Special Database publications and OCR benchmark reports.

---
Source: https://www.wikiprompt.org/wiki/mnist-dataset-1998
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-09T02:01:22.484508+00:00