Is the chat template used at all when they benchmark the model?