VibrantVista/style-judge-dataset

Style Judge Dataset

A pairwise dataset for learning a continuous style-similarity function while controlling for topic, introduced in Capturing Classic Authorial Style in Long-Form Story Generation with GRPO Fine-Tuning (arXiv:2512.05747).

Dataset Summary

• —Total rows: 156k (default subset)
• —Splits: train 130k, validation 13k, test 13k
• —Format: Arrow

Columns

• —sentence1 (string): original chunk text
• —sentence2 (string): refilled chunk text
• —score (float): calibrated similarity score in [0, 1]
• —subject (string): one of 4 subject categories
• —author1, author2 (string): author identifiers for each side
• —book1, book2 (string): book identifiers for each side
• —id1, id2 (string): example ids

Source Data (Books)

General dataset size: 978 books (strict pairs only)

Subject breakdown:

Loading

from datasets import load_dataset

ds = load_dataset("VibrantVista/style-judge-dataset")
train = ds["train"]
valid = ds["validation"]
test  = ds["test"]

Citation

@misc{liu2025capturingclassicauthorialstyle,
      title={Capturing Classic Authorial Style in Long-Form Story Generation with GRPO Fine-Tuning}, 
      author={Jinlong Liu and Mohammed Bahja and Venelin Kovatchev and Mark Lee},
      year={2025},
      eprint={2512.05747},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.05747}, 
}