Generate highly detailed plot synopses for a nearly infinite array of stories
- Keywords >> Synopsis
- Synopsis >> Plot
- Plot >> Scenes
- Scenes >> Script
- Script >> Prose
- Ensure all synopses have names, dates, places, etc. No half-assed, generic summaries!
- Use a rubric grading scheme for automatic dataset augmentation
- Start with a bunch of variables and a good prompt
- Generate a bunch of synopsis data
- Filter out bad ones (too short, too long)
- Generate many more synopses with the finetune model later
- Reverse engineer the original prompt (just a sentence or two)
- Train the model to generate a fully fledged synopsis from a small amount of inspiration
- Generate more synopses with the finetuned model and user data
- Create a feedback loop to improve the synopsis generating dataset
End goal should be a finetune dataset with the following characteristics
- 1000 samples or so
- Diverse kinds of input (different formats, structures, keywords, appeal terms, varying levels of detail)
- Consistent output (well fleshed out synopsis that tells the whole story in one big paragraph)
- Finetuned model that reliably generates top notch synopses
- List of appeal terms
- Chat log (between an author and a reader, or a librarian and patron) aka REFERENCE INTERVIEW
- "I want a book that..."
- List of comp titles
- Must have all 3 pillars: Plot, Character, Setting
- All characters must be named and have arcs
- Setting must be named, dated, described, etc
- Plot needs clear beginning, middle, and end - with narrative progression