philippkolbe/mnist-curated-with-idk-labels

Juypter Notebook with which this dataset was generated.

I used CLIP embeddings, active learning with a LeNet and mislabeled, hardness, and mistakeness metrics to manually find outliers in the MNIST dataset using fiftyone.

Numbers that are ambigous (e.g. could be either a "1" or a "7") or are fully unrecognizable are labeled as "idk".