The idea was to perform NLP like pre-processing to form the input vectors. So some tinkering involved, but the goal was to feed the network a vector of ‘word’ indexes, in this case mapping chars to an index. There are likely to be many different special characters in the log content, so choosing this method kept the number of params low (only ~90 possible chars in the data). If I were to take this further, I’d break out the structured data to be analyzed in different methods.
The dataset was only split out into training & cross validation (As seen in the Keras results, acc & loss are outputted for both). In a more formal setting I would have definitely introduced a test set as well.