Fine tuning

[zenith salmon]

This is all just a prompt engineering problem. You’re wasting money on fine tuning when you could use that for input/output tokens instead

“Makes the story not good” is subjective and you’d need to create evals for this

Changing the number of output tokens doesn’t affect the quality per se

The easiest thing to here is to have the model plan out chapters in the story, then prompt it to write each chapter

Maybe summarize the previously written chapter and include the last few paragraphs for coherence

[crimson onyx]

so i should just use gpt4o and make a prompt like:

• Please plan a first person narrative like these examples.
  then somehow i would make it like make a segment for the planned thing:
• please write a first person narrative like the examples based on this plan

[zenith salmon]

I wouldn’t dump full story examples, else it will basically copy that story

[crimson onyx]

what would you suggest then? I have a lot of examples so i hope itll just like make something similar to all of them

[zenith salmon]

You can add it to dataset for evaluation

[crimson onyx]

what's that mean