This repository contains the CIFAR-10.2 dataset prepared and published in "Harder or Different? A Closer Look at Distribution Shift in Dataset Reproduction" by Shangyun Lu, Bradley Nott, Aaron Olson, Alberto Todeschini, Puya Vahabi, Yair Carmon and Ludwig Schmidt.

The files cifar102_train.npz and cifar102_test.npz contain the train and test sets of CIFAR-10.2. Each file contains the the following keys:

• images: an nx32x32x10 numpy uint8 array containing the image data, where n=10000 for the training set and n=2000 for the test set
• labels: a numpy int64 array containing labels (integers between 0 and 9)
• label_names: a list mapping CIFAR-10 label indices to their meaning
• ti_indices: an array of the image indices in the 80 Million Tiny Images
• keywords: a list of the TinyImages keyword for every image