Sanitizing dataset with offensive content? | OpenAI | Page 1

woven mortar Dec 6, 2022, 8:52 PM

#

I'm in the process of procuring data to fine-tune GPT-3, but I noticed some of the prompts in my dataset of around 17k samples contained possibly offensive content. My first inclination would be to cut down on the number of samples so that I could run them all through OpenAI's moderation system before fine tuning.

My question is: is this realistic? Or would it be okay to leave the data as-is and moderate outputs? I am planning to train on Discord chat logs if that's relevant.

eager berry Dec 6, 2022, 9:19 PM

#

would filtering out a set of banned words not work?

woven mortar Dec 6, 2022, 9:22 PM

#

eager berry would filtering out a set of banned words not work?

That would probably work. I am worried some things will slip through the cracks but it has to be an improvement.

woven mortar Dec 6, 2022, 9:49 PM

#

Seems like it worked, I haven't found any objectionable content yet.

#Sanitizing dataset with offensive content?