How to finetune GPT3 to sound like my funny uncle?

Hi Everyone, I’m very new to GPT3, but after watching @daveshapautomator youtube channel on finetuning. I feel like I want to try my hand on fine-tuning my own model.

I have this Idea I want to try to create, the funny uncle everyone has, but have him help rewrite my own writing and emails. I’m aiming at a very specific style and tone of voice which is a bit hit and miss with prompt engineering.

I have collected a bunch of text that falls within the Tone of voice and writing style I’m looking for. They’re all of varying lengths and coming from multiple sources. I’m just a bit unsure how to approach the construction of the prompt in my training set.

{"prompt": "<prompt text>", "completion": "<ideal generated text>"}

I assume the completion should be the texts I’ve collected, but how would I go about getting a prompt?
Should I literally just go with some prompt engineering try to reverse engineer my sample texts to ‘non funny’ and use those as inputs?

Sorry if this seems very obvious to someone, I’ve been trying to figure out how I would go about this, but sofar have run into dead ends. Sorry if this is super obvious to someone, I’m just trying to learn

1 Like

This is the million dollar question I have too, but have yet to really sit down and experiment with it.

But to extract a personality, I am thinking of embeddings or fine-tunes.

For fine-tunes, see my babbling rant at Extracting Personalities from past Conversations?

For embeddings, I haven’t posted much about it, but I am thinking of the standard Q&A approach found here, and ask GPT-3 to summarize in some way that will preserve the tone. Where the data embedded is a bunch of things that person would say/think/believe.

2 Likes

Thank you for this @curt.kennedy. For a second I thought i might be the only one sitting silently with this question. Reading through the extracting personalities from past conversations. I might scale back my project.

This might be completely obvious to everyone who knows more about GPT than me. But say I found a bunch of product descriptions from the Palace skateboards webshop and trained a model on that.
Fx
Input

Product description of brown shopping bag in corduroy

Output

CORDUROY SHOPPER BROWN
I FEEL LIKE PEOPLE CALLED ROY
DON’T REALLY MESS ABOUT WITH CORDUROY

Then :crossed_fingers: I assume the model would be good at making similar product descriptions, but would it talk in a similar way about anything else?