Hacker News Reader

Do you have any explanations on why this performed better than Llama 2?

They did several things differently from Llama 2.

From my understanding, you'd have to repeat the experiment isolating each variable to see what difference each one actually makes, no?