I’ve been trying to generate stories with the davinci model, but the stories generated are quite short in length. I’ve been trying different things, from “write 500 words” to “write 3 pages” to “write 3 paragraphs” but nothing seems to really work. The length of the story created doesn’t seem to change much.
The GPT models don’t have a good sense of what “words” are, they think in tokens.
Likewise, things like writing
x number of paragraphs or pages is something GPT models are going to struggle with because it requires forethought to plan or how to stretch the information out to whatever length you want.
Which isn’t to say there’s nothing you can do.
But first you need to understand and accept you’re unlikely to be able to get it to write as much as you want in a single prompt.
With that out of the way, probably the best and most reliable way to get longer coherent results is to step in yourself and help the model where it is week. Explicitly help it plan what it’s going to write, give it a topic and maybe a few key points and ask it to write a detailed outline of what you want it to write.
Then, if you want it longer, have it go through each outline point and flush them out a little bit. Iterate through this as much as you want/need.
Finally, I would take the final outline along with the original prompt and in a new session with cleared context instruct it to complete the prompt using the outline as reference.
Depending on how long you’re trying to get, you might need to do it a little more piecemeal.
Alternately, you might try just asking it to write your story, then iteratively asking it to expand on parts.
You might try to give it an example with 1000 words maybe and ask for a text of similar length. I didn’t try that yet, but might work.
I’ve been toying asking for “an article” instead of a story, and I think it works better.
You could always try asking for the story by the fiction length classification.
Different people put different numbers to them, but here is one breakdown I found online.
I’d be interested to see what the effects are between the different prompts,
Write a short story about X
Write a novelette about X
Write a novella about X
Write a novel about X
I don’t have time to try it right now, but if anyone wants to give it a shot, I would be extremely interested to see the results.
If I remember tonight, maybe I’ll try it myself.
Sounds good, but because of token limits that can’t work yet.
Using ChatGPT 3.5 I often encountered the same issue with single replies being 500-600 words each. Not sure if that’s maybe hitting the maximum reply length.
Ultimately I used a prompt guiding the model to create a almost crazy long text but needed to prompt “continue” in order to actually get the whole text. Here is a link: continue long text
Don’t mind the example as this is just a good example regarding a topic with much to say about making it easy for the model to actually create a lengthy reply.
Here is the prompt: “Hi! Write a long and laborious text about water fasting for a prolonged period of time. Split the text into an overview and several sub paragraphs. Then summarize each paragraph and provide an overview of the details included in the paragraph. Then elaborate every detail by providing a overview and detailed deep dive before summarizing extensively. Then summarize the paragraph in great length. Finally extend on all paragraphs by creating an exhaustive and complete overview of everything written before.”
Hope this helps.