Create README.md

# Curation of the MNIST Dataset

The MNIST curation was performed using a combination of qualitative exploration and quantitative scoring.
First, visualization techniques such as PCA and UMAP were applied to the model’s embedding space in FiftyOne, enabling the identification of outliers and ambiguous digit shapes.

In addition, several FiftyOne Brain metrics were computed, including hardness, mistakenness, and uniqueness, to highlight samples with unusual characteristics or a high likelihood of misclassification. These qualitative and quantitative signals were then reviewed to curate a subset of questionable digits that were relabeled as IDK (“I Don’t Know”).

All code used for the curation process, dataset preparation, and the training of both the baseline classifier and the extended 11-class classifier (digits 0–9 + IDK) is available in the public GitHub repository:

👉 https://github.com/Conscht/MNIST_Curation_Repo/tree/main

README.md

@@ -0,0 +1,7 @@
+# Curation of the famous MNIST Dataset
+
+The curation was done using qualitative analysis of the dataset, following visualization techniques like PCA and UMAP 
+and score based categorization of the samples using metrics like hardness, mistakeness or uniqueness.
+
+The code of the curation can be found on Github: https://github.com/Conscht/MNIST_Curation_Repo/tree/main
+