After one day of writing a preprocessor and salvaging code from previous linear model, the transformer is finally training.

ETA 40minutes