While going through the semi-supervised learning example in the unsupervised learning section, I initially got a slightly different accuracy when I manually labeled the representative digits based on what I actually saw in the images.

I initially thought it was an issue in my code, but to verify, I ran the original notebook as-is on Google Colab — and to my surprise, the model's accuracy was just 7%.

After digging into the code, I found the problem:

The hardcoded labels for the 50 representative digits (y_representative_digits) no longer match the current cluster centroids generated by KMeans. This is likely due to internal changes in the dataset order or scikit-learn's clustering behavior (like randomness in centroid initialization or data shuffling).

Because of this mismatch, the model was being trained on incorrect image-label pairs, leading to terrible accuracy.

Fix:

Replaced outdated y_representative_digits with correct labels (manually reassigned by inspecting the actual centroids).

Note:

I also have an earlier PR open
#196
Kindly review that one as well.