Adjusting the dataset

Hello!
I am really happy with the project that you have put forward. However, there is a problem that I could use your help in fixing with the dataset.
The dataset includes many letters/numbers that are similar to one another. What I mean is, a handwritten be is often mis-interpreted as a 6 - even by humans. There are many such letters, which I would like to take out of the dataset. I kindof did this by doing !rm -rf {directory} on the images that I didn't want to include, however the model still predicts them nonetheless. How can I circumvent this?

Edit: I'm sure it has something to do with editing the CSV, and I'm currently trying to figure this stuff out via your dataprocessing.ipynb file, but i'm struggling

Start with dataprocessing.ipynb and inspect how the CSV maps image directories to labels, then compare those entries with the deleted images. Trace how the resulting dataset is used for model prediction; done means the excluded letter/number classes are no longer represented in the processed dataset or predictions.

Written by the indexing model from the issue text.