Comment by 20after4

I played with llama2 for a bit to see if this was easy to recreate. It's surprisingly easy to get 90% of the way there but I have yet to get the results to be as appropriate. I wonder if there is some fine tuning or other forms of extended context (beyond a system prompt) to make this magic happen. Seems like most likely it's not just a vanilla llama2 model + a bit of prompt engineering..