LLaMa is kinda insane

[Mr-NI]

is this normal?

[Green-Sky]

Yes, llama was not finetuned and contains all kinds of hostile and nsfw content scraped from the web.
It also depends on what your initial prompt was.

[iHaagcom]

Can we fine tune it? Is so how?

[Green-Sky]

how?

this is an example: https://github.com/tatsu-lab/stanford_alpaca

[slaren]

You can run a fine tuned 7B model with llama.cpp already using https://github.com/tloen/alpaca-lora. Follow the install instructions and run export_state_dict_checkpoint.py, you will get a .pth that can be converted and quantized with llama.cpp.