I have been working on a horror novel simply by prompting GPT4 and collecting results into some pretty nifty storyline content. Sometimes id have GPT 4 play a game with me where it would be a villain character in the story talking to me, and I the hero. Now I am wondering, with the about 300+pages of rough draft, is there a way I could fine tune GPT 3.5 turbo with these many pages of text with the understanding that… I want to keep engaging it to assist me with more content generation with a story that has the kind of mood set in those first few hundred pages of novel I have generated? Im a complete amatuer with code, im somewhat familiar with python, ive done some copy paste work and mostly have GPT teach me how to use it lol. So go easy on me.
Welcome to the forum.
Yes, this is possible with fine-tuning. You’d want to chunk the content into small bites.
You can find out more here in the docs…
Thank you… is there any known users, individuals or perhaps anyone here in the forum that does this kind of thing? Not a very tech savvy individual. I JUST discovered fine tuning as I was trying to create story content. And then I look at this like some kind of “snowball effect” for accelerating what I already started. I dont know how anyone thought AI was going to “replace” authors lol, its like saying tractors replace farmers. Nope, farmers just use tractors to magnify their will.
I’ve not fine-tuned recently, but I fine-tuned more than one “fiction datasets” for GPT-2 back in the day. It’s not easy, but it can be useful. You want to make sure your dataset is as clean as possible. I would leave the user blank and just use assistant probably… or maybe mix half in user and the rest of the chunk in the assistant as the “output” for the example.
The other thing you can do (if you have access) is to use GPT-3.5 16k or GPT-4 32k to just feed some of the context of your novel to the model. A lot faster, but it can get expensive.
And no, it’s not a replacement for authors yet, but it’s a great tool to double or triple your hourly output of words.
I’ve built lot of my own tools. This is more for short-form than long-form, but it’s great for brainstorming, characters, etc…
So I guess its pretty straight forward. I guess you just use python and your OpenAI API key, pay the bill and stuff and you got a fine tuned model. Ive done this with docker/autogpt, which I wasnt really impressed with. So its about as easy as that I imagine? Your just running your novel rough draft thru it and training it? I was reading in the official docs it takes time you get an email when its done and that you pay for a certain number of “epochs” or something?
Well, you need to “clean” the data, as I mentioned, but yeah, pretty straight-forward.
Stop back by if you get stuck and have a question… The big thing is going to be how you set up your dataset. You can’t just simply feed it your novel.txt…