README.md

metadata

license: mit
dataset_info:
  - config_name: default
    features:
      - name: text
        dtype: string
      - name: entity1
        dtype: string
      - name: entity2
        dtype: string
      - name: relation
        dtype: string
    splits:
      - name: train
        num_bytes: 279992108
        num_examples: 1354667
      - name: test
        num_bytes: 36572953
        num_examples: 178866
    download_size: 208297378
    dataset_size: 316565061
  - config_name: subsampled
    features:
      - name: text
        dtype: string
      - name: entity1
        dtype: string
      - name: entity2
        dtype: string
      - name: relation
        dtype: string
      - name: prompt_0_shot
        dtype: string
      - name: prompt_2_shot
        dtype: string
      - name: prompt_5_shot
        dtype: string
    splits:
      - name: train
        num_bytes: 775716423
        num_examples: 238508
      - name: validation
        num_bytes: 97227971
        num_examples: 30000
      - name: test
        num_bytes: 102379821
        num_examples: 31492
    download_size: 461712759
    dataset_size: 975324215
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
  - config_name: subsampled
    data_files:
      - split: train
        path: subsampled/train-*
      - split: validation
        path: subsampled/validation-*
      - split: test
        path: subsampled/test-*
task_categories:
  - feature-extraction
  - text-generation
  - text-classification
language:
  - en
tags:
  - biographies
  - relation-extraction
  - information-extraction
size_categories:
  - 1M<n<10M

Biographical Dataset for Relation Extraction (RE)

Overview

This dataset is a reconstructed version of the Biographical Dataset, specifically designed for relation extraction (RE) tasks. It serves as a valuable resource for digital humanities (DH) and historical research, enabling the study of relationships within biographical data. The dataset is generated by automatically aligning sentences from Wikipedia articles with structured data sourced from platforms like Pantheon and Wikidata. It includes the following key features:

• text: The textual sentence containing the entities and their relationship.
• entity1: The first entity in the relationship.
• entity2: The second entity in the relationship.
• entity1Type: The type or category of the first entity.
• entity2Type: The type or category of the second entity.
• relation: The labeled relationship between the entities.

This copy is packaged for 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. The subsampled config additionally carries ready-to-use instruction prompts (prompt_0_shot, prompt_2_shot, prompt_5_shot) for prompt-conditioned fine-tuning and evaluation of (small) language models.

Dataset Highlights

• Semi-Supervised Approach: Combines textual and structured data using a semi-supervised alignment method to ensure both scalability and accuracy.
• Tailored for Digital Humanities: Designed to assist DH researchers in analyzing historical and biographical relationships.
• Multi-Source Compilation: Integrates data from Wikipedia, Pantheon, and Wikidata to ensure a rich and reliable dataset.
• Optimized for Relation Extraction (RE): Provides labeled examples for training and evaluating models that extract relationships from unstructured text.

Configs

• default — the full reconstructed dataset, fields: text, entity1, entity2, relation (splits: train, test).
• subsampled — a smaller subset used in the paper, with the same fields plus the prompt_0_shot / prompt_2_shot / prompt_5_shot instruction-prompt columns (splits: train, validation, test).

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.

Potential Applications

• Machine Learning in RE: Training and testing models for extracting relationships from textual data.
• Historical Research: Investigating biographical and historical connections for digital humanities projects.
• Knowledge Graph Enhancement: Adding reliable relationship data to knowledge graphs from biographical information.

How to use

The Biographical Dataset is accessible through the HuggingFace Datasets library. Here's a sample code snippet:

from datasets import load_dataset

# Load the full dataset
biographical = load_dataset("Despina/biographical")

# ...or the subsampled config used in the paper (with prompt columns)
biographical_sub = load_dataset("Despina/biographical", "subsampled")

# Inspect the dataset
print(biographical["train"][0])

Data Sources

• Wikipedia: Extracted sentences from biographical articles.
• Pantheon: Structured data on notable historical figures.
• Wikidata: Comprehensive structured knowledge repository.

License

This dataset is distributed under the MIT License.

Citation

If you use this dataset, please cite our paper along with the original Biographical dataset 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}
}

@inproceedings{plum2022biographical,
  title={Biographical: A Semi-Supervised Relation Extraction Dataset},
  author={Plum, Alistair and Ranasinghe, Tharindu and Jones, Spencer and Or{\u{a}}san, Constantin and Mitkov, Ruslan},
  booktitle={Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  pages={3121--3130},
  year={2022}
}