Source-linked AI summary
Scaling Creative Writing Beyond Story-Centric Data with Attribute-Guided Genre Expansion
Hwan Chang, Yongil Kim, Heuiyeen Yeen, Yireun Kim, Jinsik Lee, Hwanhee Lee
TL;DR
Story-centric creative-writing data does not adequately represent the structural conventions of diverse formats. This paper introduces attribute-guided genre expansion and finds that its 50K-example, 13-genre dataset improves creative-writing performance and output novelty.
Problem
Existing creative-writing resources are heavily story-centric, despite diverse formats having distinct structural, stylistic, functional, and formatting conventions.
Method
The framework expands human-authored story prompts into genre-faithful instruction data using manually curated attributes that control genre-specific conventions.
Results
Models fine-tuned on the Multi-Genre Collection outperform base models, writing-specialized baselines, and existing writing-dataset models across benchmarks, while novelty rises with genre coverage.
Takeaways & Limitations
Controlled expansion across 13 genres provides more transferable creative-writing training data and enhances output novelty.
Abstract
from arXiv · showhide
High-quality creative writing data for large language models (LLMs) remains dominated by story-centric data, limiting models' ability to follow the structural and functional conventions of diverse creative formats. We propose an attribute-guided genre expansion framework for scaling creative writing data beyond story generation. By separating thematic breadth from genre-form control, our framework leverages human-authored story prompts as diverse creative seeds, while utilizing manually curated genre attributes to enforce distinct structural, stylistic, and formatting conventions. We combine these to prompt strong LLMs for genre-faithful query-response pairs, which are then quality-filtered. Applying this framework, we construct the Multi-Genre Collection, a 50K-example corpus spanning 13 creative genres, including story, rap, lyrics, scripts, game design, character design, and other creative formats. Experiments across out-of-distribution writing benchmarks and held-out genre diagnostics demonstrate that models fine-tuned on our data consistently surpass not only base models and writing-specialized baselines, but also models trained on existing writing corpora. Genre-count ablations further indicate that controlled genre expansion, rather than story-centric scaling alone, is a key driver of robust creative writing capability.
1 Introduction
Creative-writing resources remain heavily story-centric, leaving non-story genres underrepresented and their structural conventions insufficiently encoded. The paper addresses this gap with attribute-guided genre expansion, a 50K-instance corpus spanning 13 genres that improves creative-writing capability across benchmarks.
- Motivation: Public creative-writing resources are dominated by story-centric narrative generation, despite users requesting formats such as rap verses and game design documents.These genres differ beyond topic, requiring distinct formal conventions.
- Motivation: Scaling story-centric data alone does not systematically encode the formal constraints required by non-story genres.Generic synthesis can produce topically diverse but formally under-specified prompts, motivating human guidance for genre-form constraints.
- Approach: The framework separates thematic variation from genre-form control by combining diverse human-authored story prompts with manually curated genre attributes.It targets systematic genre coverage, genre-form fidelity, and specification diversity.
- Contribution: 50K instances span 13 creative-writing genres in the resulting Multi-Genre Collection.The collection is designed to scale creative-writing data beyond story-centric sources.
- Results: Models fine-tuned on the Multi-Genre Collection consistently outperform base models, LongWriter-glm4-9B, and models trained on existing writing corpora across popular benchmarks.Increasing genre coverage also improves output novelty relative to story-only training.
2 Attribute-Guided Genre Expansion
The Multi-Genre Collection is constructed through attribute-guided genre expansion, combining human-authored story seeds with genre-specific attributes to generate genre-faithful creative writing pairs. A three-stage pipeline produces a 50K-instance English corpus spanning 13 creative genres, with quality filtering and balanced genre coverage.
- Framework: The framework expands beyond story-centric datasets by generating queries for each target genre through attribute-guided genre expansion.It starts from a human-authored story-centric seed pool rather than directly collecting prompts for every genre.
- Framework: The pipeline separates thematic breadth from genre-form control: seed examples provide topical and stylistic variation, while attributes enforce structural conventions.Genre attributes include dimensions such as rap rhyme scheme and flow, or TV-series narrative arc and character development.
- Pipeline: Three stages sample n=5 thematic seed pairs, sample genre attributes, and synthesize attribute-guided query-response pairs with LLMs.The synthesis stage generates five queries per template instantiation, uses verbalized sampling for diversity, and filters low-quality responses with an independent LLM judge.
- Outcomes: The resulting Multi-Genre Collection contains 50K instances with balanced coverage across 13 creative genres.The “Etc.” category includes long-tail formats such as diary entries, comic scripts, and interactive fiction; t-SNE visualization is used to assess semantic distinctiveness.
3 Experiments
Experiments show that SFT on the Multi-Genre Collection improves base models across out-of-distribution and in-distribution creative-writing evaluations. Gains remain robust against independent judges and human evaluation, while genre diversity improves novelty and format-specific writing quality.
- Models and Benchmarks: Three base models are fine-tuned with LoRA-based SFT and evaluated on Arena Hard, WritingBench, and a held-out Multi-Genre test set.The models are Llama-3.1-8B-Instruct, EXAONE-3.5-7.8B-Instruct, and Qwen3-8B.
- Overall Performance: All three base models improve substantially after SFT, with consistent out-of-distribution gains and balanced coverage across 13 genres.The largest out-of-distribution gains occur for Qwen3-8B.
- Overall Performance: All fine-tuned models outperform the writing-specialized LongWriter-glm4-9B baseline, showing the benefit of diverse genre-specific data over narrow specialization.This comparison is reported as a consistent result across the evaluated models.
- Comparison with Existing Writing Datasets: Qwen3-8B trained on the Multi-Genre Collection substantially outperforms models trained on DeepWriting-20k and LongWriter-6k across all benchmarks.The comparison uses randomly sampled 2K examples from each dataset for SFT, with particularly large gains on Arena Hard.
- Impact of Genre Diversity: NoveltyBench Distinct scores rise monotonically as training genres increase from 0 to 13, indicating that genre diversity enhances output distinctness.The metric estimates distinctness by clustering semantically and functionally equivalent generations.
- Robustness and Qualitative Analysis: Independent-judge evaluation and human evaluation confirm that fine-tuned variants outperform their base models, while qualitative outputs show improved format compliance, specificity, and coherence.The qualitative example adds titles, named entities, a recurring character, specific spells, and section-level timestamps.
4 Related Work
Prior creative-writing datasets have largely centered on story generation, with established resources covering prompt–story pairs, short commonsense narratives, and preference-based evaluation. Other work has addressed individual genres separately rather than broad genre coverage.
- 4 Related Work: Early creative-writing datasets focused almost exclusively on story generation.This included resources designed for narrative understanding and story-quality evaluation.
- 4 Related Work: WritingPrompts provided diverse prompt–story pairs sourced from Reddit.It was introduced as a dataset for creative writing research.
- 4 Related Work: ROCStories offered five-sentence commonsense stories for narrative understanding.The dataset targeted concise narrative examples rather than multiple creative genres.
- 4 Related Work: LitBench introduced preference labels for evaluating story quality and writing personalization.Its evaluation focus remained centered on stories and writer-specific preferences.
- 4 Related Work: Prior work beyond stories targeted individual genres in isolation.The passage begins an example with SongCompos, but does not provide further details.
5 Conclusion
The paper introduces the 50K-instance Multi-Genre Collection spanning 13 genres, built through attribute-guided genre expansion. Experiments show that models trained on it outperform base models, writing-specialized baselines, and existing writing datasets.
- Dataset: The Multi-Genre Collection contains 50K instances across 13 genres.It is the paper’s introduced dataset.
- Method: Attribute-guided genre expansion transfers thematic diversity from human-authored prompts while enforcing genre-specific constraints with manually curated attributes.The framework combines diverse thematic seeds with explicit genre control.
- Findings: Models fine-tuned on the collection substantially outperform both base models and writing-specialized baselines.This is one of the paper’s three reported experimental findings.
- Findings: The collection consistently outperforms existing writing datasets.This is another of the paper’s three reported experimental findings.
GenAI Usage Disclosure
The manuscript was written by the authors, with ChatGPT used only to refine style, clarity, and grammar. Large language models also serve as components of the data-construction pipeline for query generation, response generation, filtering, and quality scoring.
- ChatGPT was used solely to refine the manuscript’s style, clarity, and grammar, not for ideation, claim generation, or experimental design.
- The data-construction pipeline uses GPT-5-mini for filtering and query generation, Qwen3-235B-A22B-Thinking for response generation, and Qwen3-30B-A3B-Instruct as a quality-scoring LLM-as-a-judge.