Create a model style for open-llama-open-instruct

OoriData / OgbujiPT

Client-side toolkit for using large language models, including where self-hosted

Apache License 2.0

103 stars 8 forks source link

Create a model style for open-llama-open-instruct #5

Closed uogbuji closed 1 year ago

uogbuji commented 1 year ago

We installed TheBloke/open-llama-7b-open-instruct-GPTQ on our Ooba (see Reddit thread). It seems you can prompt it with a direct prompt, or alpaca-instruct style, but I'm confirming this on Ooba Discord.

@choccccy, useful task for you. We might well actually just also put it into alpaca.py and rename that file llama_alpaca.py (LLaMA comes before Alpaca in LLM lineage).

uogbuji commented 1 year ago

Some notes I found on Reddit; nothing on straight LLaM, though.

vicuna (used by e.g. stable vicuna)

### Human: {{prompt}}
### Assistant:{{gen}}

wizard (used by e.g. wizard vicuna)

USER: {{prompt}}
ASSISTANT:{{gen}}

alpaca (used by e.g. gpt4-x-vicuna)

### Instruction: {{prompt}}
### Input: {{input}}
### Response:{{gen}}

choccccy commented 1 year ago

We will probably eventually have more modular and more universal prompting templates, but for now we have most of the usual gamut of prompting covered already.