ch-chat-70b

Experimental merge to extract the good stuff and remove the bad stuff from llama 2 chat, should be good for chat/roleplay and other creative usecases.

In general, it seems to keep the system prompt adherence and prose qualities of Llama 2 chat, but without some of the negative aspects like being overly-repetitious or formulaic.

This was really just a test model, but I'm pleased enough with the results to release it. I'll work on more polished merges soon since this recipe can almost certainly be refined.

Don't expect it to be totally uncensored, it still has enough llama 2 chat in it that it'll drop a refusal on you occasionally if your context is empty. Using a system message guiding it towards your desired content will greatly reduce this behavior.

Prompt Format

I recommend using Alpaca with Llama-2-chat style system prompt, with or without the Input tag depending on your preferred prompt style:

<<SYS>> <system message/character card/whatever>
<</SYS>>

### Instruction:
<user input/instruction>

### Response:
<model response>

<<SYS>> <system message/character card/whatever>
<</SYS>>

### Instruction:
<instruction>

### Input:
<information/history>

### Response:
<model response>

Regular Alpaca works too, but it's better with a system prompt. Handles both single-instruction and multi-turn well in my experience.