Scaling Creative Writing Beyond Story-Centric Data with Attribute-Guided Genre Expansion

The Problem: LLMs Trained Only on Stories Can't Write in Other Forms

Current large language models excel at story generation because nearly all their creative writing training data comes from stories. When asked to write a rap, screenplay, or game design document, they struggle—not because they lack language ability, but because they've never learned the specific structural and stylistic rules that govern these formats. The authors argue that scaling creative writing training requires moving beyond story-centric data and building datasets that span diverse genres with their own distinct conventions.

The Solution: Attribute-Guided Genre Expansion

Rather than manually writing thousands of examples in each genre, the authors propose a framework that uses existing story prompts as creative seeds, then systematically expands them into other formats. Here's how it works:

The framework starts with human-authored story prompts (which provide thematic diversity and creativity). For each genre—rap, lyrics, scripts, game design, character design, and nine others—human annotators curate a set of "genre attributes": rules describing the structure, style, formatting, and tone unique to that format. For example, a rap might require rhyme schemes and rhythm, while a game design document needs mechanics, objectives, and win conditions.

These attributes are then combined with the story prompts to create instructions for strong existing LLMs (like GPT-4). The model receives a prompt: "Generate a [genre] based on this story premise [prompt], following these genre attributes [attributes]." The model outputs a query-response pair—a user request and the creative work itself. Human annotators filter these outputs for quality, keeping only the best examples.

The result is the Multi-Genre Collection: 50,000 training examples across 13 creative genres. This isn't just stories rewritten in different voices; it's data where each genre follows its own rules.

Why It Works and Why It Matters

The key insight is separating thematic breadth (what stories are about) from genre-form control (how different formats are structured). Stories provide unlimited creative variation; genre attributes enforce the rules. This combination is far more efficient than trying to hand-author diverse examples in each format.

Models fine-tuned on this data outperform both base models and existing writing-specialized baselines on out-of-distribution benchmarks—meaning they generalize to new prompts and genres they've never seen. Ablation studies show that simply scaling story data doesn't achieve the same improvement; controlled genre expansion is what drives robust creative writing.

The practical impact is significant: as creative AI applications grow (game design tools, screenplay assistants, music generation), models need to respect genre conventions. This work demonstrates that with a modest amount of structured annotation (curating attributes for each genre), you can efficiently create training data that teaches models these conventions at scale. The 50K-example corpus is relatively small by modern standards, yet produces measurable gains, suggesting the framework scales efficiently.