Researchers from Hugging Face have introduced an attribute-guided genre expansion framework designed to scale creative writing data beyond traditional story-centric datasets. The approach addresses a key limitation in current LLM training data, which often focuses heavily on narrative forms, leaving models less adept at handling the structural and stylistic conventions of other creative formats.
The framework works by separating thematic breadth from genre-form control. Human-authored story prompts serve as diverse creative seeds, while manually curated genre attributes enforce distinct structural, stylistic, and formatting rules. These elements are combined to prompt strong LLMs to generate genre-faithful query-response pairs, which are then filtered for quality.
Using this method, the team constructed the Multi-Genre Collection, a corpus of 50,000 examples spanning 13 creative genres, including story, rap, lyrics, scripts, game design, and character design. Experiments on out-of-distribution writing benchmarks and held-out genre diagnostics showed that models fine-tuned on this data consistently outperform base models, writing-specialized baselines, and models trained on existing writing corpora.
Genre-count ablations further indicated that controlled genre expansion—rather than simply scaling story-centric data—is a key driver of robust creative writing capability. The findings suggest that diversifying training data across genres can significantly enhance LLMs' creative writing abilities.