TheBloke

/

CodeLlama-70B-Python-GGUF

like 39

Same here. Moved on. ( and the size was killing my machine too, but for no benefit )

You have to use the instruct codellama one. This is not for instruction but autocomplete.

Base mixtral 8x7b is also autocomplete but the instruct mixtral is probably what u are talking about

Instruct gguf made by Thebloke:(https://huggingface.co/TheBloke/CodeLlama-70B-Instruct-GGUF)

Instruct gguf made by Thebloke:(https://huggingface.co/TheBloke/CodeLlama-70B-Instruct-GGUF)

Ill try that before i give up on it. i may have just missed the instruct version.

Also tested instruction version and the same poor results.

Also tested instruction version and the same poor results.
Ya, i was just told "as an AI bla bla bla it is not appropriate or ethical to provide you python code to add 2 numbers.. bla bla bla bla bla bla "

yep that model is as bad as 34b llama version ... or surprise me even worse .. mixtral or wizard coder 1.1 are doing much better job right now.

Did you follow the correct prompt format?

Also tested instruction version and the same poor results.
Ya, i was just told "as an AI bla bla bla it is not appropriate or ethical to provide you python code to add 2 numbers.. bla bla bla bla bla bla "

lol if that’s true with correct prompt format, the model is probably the most censored one ever.

I do understand that censoring llm at all. If you want something "illegal" you will find it in the internet very easily ... that is so stupid.

I do understand that censoring llm at all. If you want something "illegal" you will find it in the internet very easily ... that is so stupid.

Agreed, but its the world we live in today :(

Also tested instruction version and the same poor results.

off topic a bit:

1. Is it possible to run Q4_K_M with rtx2070(8GB), ram 48GB? How slow?

2. What coding question would you use to test LLM?

I guess time series models are illegal now according to this woke piece of junk, what a joke.