Based on a book by Tariq Rashid "Neurale Netzwerke selbst programmieren" (Übersetzung Frank Langenau, O'Reilly).
Their Github can be found here.
The model is first trained with a dataset of 60'000 samples of labeled, handwritten digits. After that, it is tested with a labeled set containing 10'000 unseen samples. Some falsely identified digits are displayed and a performance rating is given. The trained network can also be stored in a file to save time and to be able to experiment with different testing data later in time.
Typical accuracy is up to 97.5 %. The data set is not included but can be found here: Training data, Testing data