Comment by sunandcoffee

Looks like the model is slightly worse on language ability but better on coding ability. This might be a good model for trying agents.