CodeLlama-70B-Instruct

#2
by eramax - opened

CodeLlama-70B-Instruct please

ok, I can try but I'm not sure if the calibration data I use works best for that. Maybe one should make a special one for code models. I'm just gonna try it with the one I have.

KnutJaegersberg changed discussion status to closed

I'm uploading the result, but on my first two request, it told me it's unethical. I would suggest the hypothesis that that comes from the calibration data only being words, not code. So far I don't know what calibration data is good for a code model.

taking a look at https://huggingface.co/TheBloke/CodeLlama-70B-hf-GPTQ let's me think I could retry it with evol instruct code data.

i've aborted the upload will retry with that data. might take a little while.

I've uploaded a version where I used evol instruct code. I don't think it's a great model, maybe it's not even useful. I feel making this work for code models is harder than for language.

Sign up or log in to comment