Finetune on "uncensored" dataset?

#32
by sivarajan - opened

The datasets used for fine-tuning the model introduce significant bias in responses, and marked reduction in capability, famously with the verbal tic "I'm sorry, but as a large language model … ". Have you considered finetuning Falcon on datasets with such responses removed?
See evol_instruct_unfiltered and ShareGPT_unfiltered.

That would be amazing!

The censored models are not only biased, but as a result less useful.

Sign up or log in to comment