A transformer-based numerical language model on reverse-Polish Notation arithmetic
train.py: trains the model, logging toruns/<timestamp>/eval.py: evaluation over all checkpoints, writeseval_log.csvsrc/: core code (data.py,model.py,train_utils.py,eval_utils.py).plot.py: Plotting.
Addition is noticeably easier relative to multiplication.
Accuracy vs. step for addition tasks:
Accuracy vs. step for multiplication tasks:

