Despina/project-gutenberg-fiction-relations

Project Gutenberg Fiction Relations

A literary-domain relation extraction (RE) dataset built from public-domain fiction in Project Gutenberg. Each example pairs a passage of narrative text (mentioning a head and tail entity) with the relation that holds between the two entities, providing an RE resource for literary and digital-humanities research where general-domain (news / Wikipedia) datasets do not transfer well.

This dataset is released as part of the paper "Sub-Billion, Super-Frontier: Fine-Tuned Small Language Models Rival Zero-Shot Frontier LLMs on General and Literary Relation Extraction" (Christou & Tsoumakas, 2026) arXiv:2606.22606. Rows are pre-formatted as instruction prompts so the dataset can be used directly for prompt-conditioned fine-tuning and evaluation.

Annotation

Relations were automatically annotated by a GPT-4-class model (see the paper for the exact prompt and procedure); the labels are therefore silver-standard rather than human gold, and models tuned on this data partly learn the annotator's label distribution. The raw relation labels were mapped onto a canonical ontology (137 → 48 relation types); that mapping is included with the dataset.

Dataset structure

Splits: train, validation, test.

Column	Description
text	The narrative passage containing the two entities.
entity1	The head entity.
entity2	The tail entity.
relation	Gold (model-annotated) relation label between entity1 and entity2.
prompt_0_shot	Zero-shot instruction prompt (task instructions + the input passage).
prompt_2_shot	The same prompt with 2 in-context demonstrations prepended.
prompt_5_shot	The same prompt with 5 in-context demonstrations prepended.

The three prompt_* columns are alternative renderings of the same example at different shot counts, so pick one shot setting per experiment rather than concatenating them. Inspect the relation column for the exact 48-relation label set.

Usage

from datasets import load_dataset

ds = load_dataset("Despina/project-gutenberg-fiction-relations")
print(ds["test"][0]["prompt_2_shot"])  # formatted input
print(ds["test"][0]["relation"])       # gold label

Source and licensing

The underlying texts come from Project Gutenberg, which distributes works that are in the public domain in the United States. Note that Project Gutenberg's own trademark and website terms apply to material redistributed under the "Project Gutenberg" name; the public-domain source texts themselves are free of copyright restrictions. The relation annotations in this dataset are contributed by the authors of the paper below.

Citation

If you use this dataset, please cite our paper:

@article{christou2026subbillion,
  title        = {Sub-Billion, Super-Frontier: Small Language Models Rival
                  Zero-Shot Frontier LLMs on General and Literary Relation Extraction},
  author       = {Christou, Despina and Tsoumakas, Grigorios},
  journal      = {arXiv preprint arXiv:2606.22606},
  year         = {2026},
  url          = {https://arxiv.org/abs/2606.22606}
}