Llama is neither instruction finetuned nor human reinforcement learning ranked. It‘s a base model like the original GPT3 that needs much more careful prompting and few-shot reasoning to produce similar results.