Comment by zinccat

My feeling is that a lot of challenge could come from the tokenizer used by the model, similar to r in strawberry problem.