Reproduce “COIG-Writer: A High-Quality Dataset for Chinese Creative Writing with Thought Processes” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yunwen Li, Shuangshuang Ying, Xingwei Qu, Xin Li, Sheng Jin, Minghao Liu, Zhoufutu Wen, Tianyu Zheng and 11 more

COIG-Writer introduces a high-quality Chinese creative writing dataset that includes explicit thought processes, demonstrating that process supervision significantly improves creative writing capabilities, achieving a 25.5 percentage point win rate increase in Chinese when balanced with general data. This work also reveals limited cross-lingual transfer of creative abilities and an inverse correlation between lexical diversity and perceived creative quality.

Key insights

  • Creative writing is compositional, consisting of distinct components: narrative logic (structural planning) provided by process supervision and linguistic expression (stylistic realization) maintained by general-purpose data.
  • Optimal creative writing performance requires a critical stabilization threshold of general-purpose data, with a ratio of approximately 1:12 (creative to general samples) achieving the best results for process supervision.
  • Creative capabilities are culturally bound and do not transfer significantly across languages, indicating the necessity of culturally and linguistically grounded data for authentic AI creativity.
  • Higher lexical diversity, as measured by Type-Token Ratio (TTR), inversely correlates with perceived creative quality, suggesting that high TTR might compensate for a lack of narrative coherence in lower-quality outputs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.