EMNIST — handwritten letters and digits, MNIST-style IDX

A verbatim mirror of the EMNIST dataset (Cohen, Afshar, Tapson & van Schaik, 2017), repackaged so each split's IDX files are fetchable individually. No pixels or labels added, removed, or reordered relative to NIST's release — only the hosting moved: the per-split members were extracted out of NIST's single bundled gzip.zip and re-uploaded at the repo root, byte-for-byte.

Re-hosted under Heliosoph for ingestion-pipeline stability. NIST ships every split inside one ~540 MB gzip.zip; a pipeline that wants just the balanced split shouldn't have to download the whole bundle and dig a member out of an archive. Pinning a HuggingFace revision lets pipelines fetch the same bytes anonymously and reproducibly, one split at a time.

Credit: Gregory Cohen, Saeed Afshar, Jonathan Tapson, André van Schaik (Western Sydney University) — glyphs drawn from NIST Special Database 19.

Why a mirror?

NIST distributes EMNIST as a single gzip.zip (~540 MB) whose members live under a gzip/ prefix. Extracting one split means downloading the entire bundle — every other split included — and then pulling a specific .gz out of the zip. Neither step is hostile, but neither plays nicely with a reproducible ingestion pipeline that just wants one split's bytes. This mirror flattens the members to the repo root so each .gz is a direct, anonymous, revision-pinned download. The bytes inside each .gz are identical to NIST's.

What this repo contains

Twelve gzip-compressed IDX files — three splits, each with a train + test cut, each cut as an images file + a labels file:

emnist-balanced-train-images-idx3-ubyte.gz   # 112,800 glyphs, 47 classes
emnist-balanced-train-labels-idx1-ubyte.gz
emnist-balanced-test-images-idx3-ubyte.gz    #  18,800 glyphs
emnist-balanced-test-labels-idx1-ubyte.gz
emnist-letters-train-images-idx3-ubyte.gz    # 124,800 glyphs, 26 classes
emnist-letters-train-labels-idx1-ubyte.gz
emnist-letters-test-images-idx3-ubyte.gz     #  20,800 glyphs
emnist-letters-test-labels-idx1-ubyte.gz
emnist-byclass-train-images-idx3-ubyte.gz    # 697,932 glyphs, 62 classes
emnist-byclass-train-labels-idx1-ubyte.gz
emnist-byclass-test-images-idx3-ubyte.gz     # 116,323 glyphs
emnist-byclass-test-labels-idx1-ubyte.gz

EMNIST upstream defines three more splits (bymerge, digits, mnist); this mirror carries the three that cover the useful range — class-balanced (balanced), pure-alphabet (letters), and the full unbalanced character set (byclass).

File format — MNIST IDX

Each split is a pair of big-endian IDX files, identical in layout to the original MNIST distribution:

File	Magic	Shape	Dtype	Meaning
*-images-idx3-ubyte	0x00000803	[N, 28, 28]	uint8	28×28 grayscale glyph, 0 (black) – 255 (white)
*-labels-idx1-ubyte	0x00000801	[N]	uint8	integer class id; vocabulary depends on the split

Row i of the images file pairs with row i of the labels file. Any MNIST loader reads these unchanged.

⚠️ Orientation — glyphs are stored transposed

EMNIST images are stored transposed relative to display orientation: the NIST → MNIST conversion baked in a 90° rotation plus a horizontal mirror. Read raw, every character looks rotated and flipped. The fix is a single matrix transpose (reflect across the main diagonal):

img = raw.reshape(28, 28).T   # upright

This is a property of the upstream bytes, not the mirror — any faithful EMNIST consumer must transpose. (DatumV applies image_transpose() in its install recipe, so the ingested images are already upright.)

Class labels

The label integer's meaning depends on the split:

• balanced (0–46): 0–9 are digits; 10–46 are letters, with the 15 upper/lowercase pairs that are visually identical (C/c, O/o, S/s, …) merged into a single class. Equal samples per class.
• byclass (0–61): 0–9 digits, 10–35 are A–Z, 36–61 are a–z. Frequencies follow natural handwriting distribution (heavily skewed).
• letters (1–26): 1–26 map to A–Z with case merged. Note the 1-based indexing — subtract 1 before indexing a zero-based class array.

How to use

Read an IDX pair directly with NumPy (no extra dependencies), remembering the transpose:

import gzip, numpy as np

def load_images(path):
    with gzip.open(path, "rb") as f:
        magic, n, rows, cols = np.frombuffer(f.read(16), dtype=">u4")
        data = np.frombuffer(f.read(), dtype=np.uint8).reshape(n, rows, cols)
    return np.transpose(data, (0, 2, 1))      # upright: transpose each glyph

def load_labels(path):
    with gzip.open(path, "rb") as f:
        magic, n = np.frombuffer(f.read(8), dtype=">u4")
        return np.frombuffer(f.read(), dtype=np.uint8)

x = load_images("emnist-balanced-train-images-idx3-ubyte.gz")  # (112800, 28, 28)
y = load_labels("emnist-balanced-train-labels-idx1-ubyte.gz")  # (112800,)
print(x.shape, y.shape, y.min(), y.max())                       # ... 0 46

Or with idx2numpy if you prefer (pip install idx2numpy):

import gzip, idx2numpy, numpy as np
with gzip.open("emnist-letters-test-images-idx3-ubyte.gz", "rb") as f:
    images = np.transpose(idx2numpy.convert_from_string(f.read()), (0, 2, 1))

Dataset specs

	Spec
Splits in this mirror	balanced (47 cls), letters (26 cls), byclass (62 cls)
Glyphs — balanced	112,800 train / 18,800 test
Glyphs — letters	124,800 train / 20,800 test
Glyphs — byclass	697,932 train / 116,323 test
Image shape	28×28, grayscale, uint8 (0–255)
Orientation	stored transposed — consumer must transpose to upright
Format	MNIST-style IDX (idx3-ubyte images, idx1-ubyte labels), gzip-compressed
Compressed size	~180 MB total across the 12 files
Language	English (Latin alphanumerics)
Source corpus	NIST Special Database 19

When to pick EMNIST

• Drop-in harder MNIST: identical 28×28 IDX shape, so any MNIST recipe runs unchanged — but strong CNNs report ~88–91% test accuracy on balanced vs. ~99.7% on MNIST. A better smoke test for whether a classifier is actually learning.
• Character / handwriting recognition: the letters split is a clean 26-class A–Z problem; byclass exercises the full 62-class digit + upper + lower vocabulary.
• Imbalanced-classification practice: byclass follows natural letter-frequency distribution — report macro-F1, not just accuracy, and reach for resampling / threshold calibration.

For plain digit classification reach for MNIST; for cursive-Japanese characters in the same IDX shape, KMNIST.

License

CC BY 4.0 (closest match; the release predates SPDX). The underlying NIST Special Database 19 is a public-domain US-government work distributed as "freely usable", and EMNIST is broadly used commercially in practice. Preserve attribution to Cohen / Afshar / Tapson / van Schaik and to NIST.

• Paper: EMNIST: an extension of MNIST to handwritten letters
• Upstream: The EMNIST Dataset — NIST
• Source corpus: NIST Special Database 19