CausalLM - STOP using wrong GPTQ model

#1
by JosephusCheung - opened

GPTQ version is not an official model, and I think it is a wrong model for its bad calibration process with wikitext dataset, as Wikipedia was trained on a new synthetic dataset, recalls based on the original text are all wrongly calibrated for GPTQ quantification.

CausalLM/14B-DPO-alpha Score : 53.77358490566038
CausalLM/7B-DPO-alpha Score : 45.59748427672956

Sign up or log in to comment