I think if we can change the llama output layer, maybe we can get logs.
j previous speech k next speech
