Dataset Overview

Dataset Description

• Dataset Name: Cifar100-LT
• Dataset Type: Image Classification
• Language: English
• License: Apache 2.0
• Dataset Size: 10K < n < 100K
• Source Dataset: cifar100
• Task Category: Image Classification
• Dataset ID: cifar-100

Dataset Summary

CIFAR‑100‑LT is an imbalanced dataset containing fewer than 60,000 colour images, each of size 32×32 pixels, distributed across 100 classes. Sample counts per class decay exponentially with factors of 10 and 100. The dataset holds 10,000 test images (100 per class) and under 50,000 training images. The 100 classes are further organized into 20 super‑classes. Each image carries two labels: a fine label for the specific class and a coarse label for the associated super‑class.

Supported Tasks and Leaderboard

• Image Classification: The goal is to assign each image to one of the 100 classes. Leaderboard available here.

Dataset Structure

Data Instances

An example from the training set:

json { "img": "<PIL.PngImagePlugin.PngImageFile image mode=RGB size=32x32 at 0x2767F58E080>", "fine_label": 19, "coarse_label": 11 }

Data Fields

• img: a PIL.Image.Image object containing a 32×32 image.
• fine_label: an int class label, e.g.:
  • 0: apple
  • 1: aquarium_fish
  • ...
  • 99: worm
• coarse_label: an int super‑class label, e.g.:
  • 0: aquatic_mammals
  • 1: fish
  • ...
  • 19: vehicles_2

Data Splits

License Information

Apache License 2.0

Citation

plaintext @TECHREPORT{Krizhevsky09learningmultiple, author = {Alex Krizhevsky}, title = {Learning multiple layers of features from tiny images}, institution = {}, year = {2009} }

Acknowledgements

Thanks to @gchhablani and all contributors who added the original balanced CIFAR‑100 dataset.