Scaling Creative Writing Beyond Story-Centric Data with Attribute-Guided Genre Expansion
Hwan Chang, Yongil Kim, Heuiyeen Yeen, Yireun Kim, Jinsik Lee, Hwanhee Lee
cs.CL
2026-08-14
An attribute-guided pipeline turns story prompts into a 50K, 13-genre corpus. Llama-3.1-8B Arena Hard writing rises from 2.9 to 33.0 after LoRA SFT.
People ask models for rap verses, screenplays, and game-design docs. Public creative-writing data is still mostly stories. Stories teach narrative. They do not teach rhyme schemes or Suno timestamps. Scaling story data leaves that structural gap intact. Code and math synthesis can lean on executability or gold answers. Creative writing has no short validity signal. Generic Self-Instruct tends to produce topically diverse prompts that underspecify form.
Three requirements follow: cover the genres users actually request, encode each genre's structural conventions, and vary how tightly the prompt constrains the writer. This CIKM 2026 paper from LG AI Research and Chung-Ang University turns those three into a repeatable synthesis pipeline.
The framework splits theme from form. Themes come from human posts on Reddit's r/WritingPrompts. GPT-5-mini removes 686 unsafe items and 861 off-topic fragments. Each generation samples five story pairs as few-shot seeds, borrowing topic, not genre.
Genre attributes are hand-curated. Encyclopedic definitions and writing manuals are scraped, GPT-5 proposes dimensions, and people merge overlaps, drop generic ones, and fill holes, leaving 5 to 15 attributes per genre. Rap gets rhyme scheme and flow; TV series get narrative arc and character development. A subset size k is drawn uniformly from zero through the full list: k = 0 is open-ended, larger k tightens the brief.
GPT-5-mini writes five queries per template call and is told to vary topic, tone, and structure. Qwen3-235B-A22B-Thinking writes the responses. A separate Qwen3-30B-A3B-Instruct judge drops pairs scoring more than two standard deviations below the mean. The result is the Multi-Genre Collection: 50K examples across 13 genres plus a long-tail Etc. bucket for diaries, comic scripts, and interactive fiction. Query embeddings form tight genre clusters under t-SNE, with little overlap.
Llama-3.1-8B-Instruct, EXAONE-3.5-7.8B, and Qwen3-8B then get LoRA SFT at rank 128 for two epochs with a 4096 cutoff. The writing-specialized baseline is LongWriter-glm4-9B.
After the full 50K SFT, all three bases rise on the Arena Hard creative-writing slice, WritingBench's creative domains, and a held-out 13-genre set with 50 items each.
| Model | Arena Hard | WritingBench | 13-genre set |
| LongWriter-glm4-9B | 2.3 | 47.2 | 49.5 |
| Llama base / +SFT | 2.9 / 33.0 | 43.7 / 60.6 | 44.4 / 68.4 |
| EXAONE base / +SFT | 10.4 / 22.9 | 52.7 / 59.4 | 64.9 / 66.2 |
| Qwen3-8B base / +SFT | 8.0 / 34.2 | 56.1 / 63.6 | 67.7 / 69.3 |
Llama jumps 30.1 Arena Hard points over its base, Qwen3 jumps 26.2, and every fine-tune beats LongWriter. A fairer 2K-per-corpus SFT on Qwen3-8B is much tighter: Multi-Genre 7.5, LongWriter-6k 7.0, DeepWriting-20k 4.9 on Arena Hard.
NoveltyBench Distinct rises monotonically from 3.87 with zero genres to 4.81 with all 13. Re-judging WritingBench with DeepSeek v3.2 and Gemini 3 Flash keeps the same ranking. On 50 human-scored prompts the scores are 59.3, 57.7, and 56.9 for Multi-Genre, LongWriter, and DeepWriting. In a qualitative Suno deathcore prompt, the fine-tuned Llama adds a title, a named world, a character, and section timestamps; the base model stays in generic chorus language.
This is an engineering recipe: human seeds for theme, hand-written attributes for form, a strong teacher for responses. If the goal is an 8B model that can write lyrics, scripts, and design docs, piling on more stories is the wrong axis. Training on 13 genres also makes outputs less redundant. The paper does not claim a new architecture.
The 2K matched comparison is only 0.5 Arena Hard points above LongWriter, and the human gap is 1.6. A large share of the 50K-versus-base jump is simply "the model got writing SFT." Genre expansion is real. It is smaller than the headline Arena Hard column suggests.
There is no dedicated limitations section. Queries and attributes come from the GPT-5 family, responses from Qwen3-235B, so the corpus inherits teacher style. The in-house test set uses that same teacher as the reference, same distribution, same taste. One primary judge, GPT-5-mini, also wrote the queries; later independent judges and a 50-prompt human sample still show a thin gap. No code or data link is provided. LoRA runs for two epochs only, with no full-SFT or preference-learning comparison. English-filtered Arena Hard and WritingBench do not speak for other languages' formal conventions.