Calibration dataset choice

#4
by iproskurina - opened

Thanks for sharing models compressed with gptq!
Why for published models wikitext is always used for calibrating when, in the GPTQ paper, authors use a sample from c4?
Does it allow for higher performance?

I believe wikitext makes the quanted model have slightly better perplexity but it’s not too much

Sign up or log in to comment