Scaling Creative Writing Beyond Story-Centric Data with Attribute-Guided Genre Expansion
The paper’s claim is that genre variety, not just more story data, improves creative-writing fine-tuning.
The authors build a 50,000-example Multi-Genre Collection across 13 formats, including rap, lyrics, scripts, game design, and character design. Their method starts from human-written story prompts, then uses curated genre attributes to force different structures, styles, and formats. After filtering the generated query-response pairs, models fine-tuned on the set beat base models, writing-specialized baselines, and models trained on existing writing corpora in the reported tests. HF Daily Papers' note
The authors build a 50,000-example Multi-Genre Collection across 13 formats, including rap, lyrics, scripts, game design, and character design. Their method starts from human-written story prompts, then uses curated genre attributes to force different structures, styles, and formats. After filtering the generated query-response pairs, models fine-tuned on the set beat base models, writing-specialized baselines, and models trained on existing writing corpora in the reported tests. HF Daily Papers' note
score 4