Comment by theshrike79

The easiest way to fix these is give the model an environment to run code.

Any model can easily one-shot a python script that can count the occurrence of any letter anywhere and return the result.

It's just a tooling issue. You really can't "train" an LLM to do it because tokenisation and ... stuff.