Part 2/8:
Shapiro begins by describing a script he crafted to synthesize detailed plot outlines tailored for different genres, settings, and periods. His method involves defining certain modifiers—for example, genre modifiers like "crime" and "mystery," locations like "America," periods such as the "1920s," and story tones like "tragic" or "heart-wrenching."
He populates a prompt template with these variables, assigning each combination a UUID to ensure diversity. Using GPT-3, he generates 396 samples, exceeding the minimum requirement of 200 samples for effective fine-tuning.
Key point: Generating more samples than needed allows for data cleansing and ensures high-quality training input.