CRD716's picture
Update README.md
120983f
metadata
license: gpl-3.0
metrics:
  - perplexity
pipeline_tag: text-generation
tags:
  - LLaMa
  - text-generation-inference
  - ggml
language:
  - en
  - bg
  - ca
  - cs
  - da
  - de
  - es
  - fr
  - hr
  - hu
  - it
  - nl
  - pl
  - pt
  - ro
  - ru
  - sl
  - sr
  - sv
  - uk

NOTE: DEPRECIATED, BETTER PEOPLE DO THIS NOW

LLaMa 65B converted to ggml via LLaMa.cpp, then quantized to 4bit.

Legacy is for llama.cpp setups older than https://github.com/ggerganov/llama.cpp/pull/1508, the regular is faster but does not work on old versions.

I recommend the following settings when running as a good starting point: main.exe -m ggml-LLaMa-65B-q4_0.bin -n -1 -t 32 -c 2048 --temp 0.7 --repeat_penalty 1.2 --mirostat 2 --interactive-first --color

Be aware that LLaMa is a text generation model, not a conversational one, and as such you will have to prompt it differently than, for example, Vicuna or ChatGPT.