Creative Writing - Part One (Shadow & Skeleton)

This dataset is designed to train Large Language Models (LLMs) in grounded creative writing by enforcing a "Think-Before-You-Write" methodology.

Methodology: Shadow & Skeleton

Most creative writing datasets train models to produce "vibes" or "cinematic descriptions" that often lack physical coherence or psychological depth. This dataset takes a different approach:

1. Shadow Prompts: We generated 1,000+ isomorphic variations of high-difficulty creative writing benchmarks. These prompts maintain the structural constraints (e.g., "Track limb position in a fight" or "Subvert romantic tropes") of rigorous benchmarks but apply them to completely new settings and characters. This prevents overfitting to specific benchmark questions while training the underlying skill.

2. Stage Blocking (The Skeleton): Every training example includes a pre-computation step (the "Stage Blocking" phase). Before writing the narrative, the model generates a logical skeleton:

   • Physical Blocking: Precise coordinate systems, weight distribution, and velocity vectors for action scenes.
   • Subtext Mapping: The hidden meaning behind dialogue lines for social scenes.
   • Cognitive Filters: The specific analytical lens of a narrator for psychological scenes.

3. Grounded Prose: The final narrative output is conditioned on this skeleton, ensuring that every description is physically possible and psychologically consistent.

Data Source

This dataset was synthesized using a high-speed reasoning model optimized for low-latency chain-of-thought generation. The pipeline used an adversarial filtering method to reject "slop" (clichés, purple prose) and prioritize concrete, sensory details.

Structure

Each entry contains:

• instruction: The creative writing prompt with specific constraints.
• output: A two-part response containing the Stage Blocking followed by the Narrative.
• metadata: Information about the abstract schema and constraints used.

Usage

This dataset is ideal for:

• Fine-tuning models to avoid "hallucinating action" (teleporting limbs, impossible physics).
• Training models to write subtext-heavy dialogue.
• Reducing "GPT-isms" and flowery verbosity in creative outputs.

Stats

Cost estimated using Unknown pricing on OpenRouter ($1.0/M input, $5.0/M output)