Despina/rebel-dataset

REBEL Processed Relation Dataset

Summary

This dataset is a processed version of the Babelscape/rebel-dataset. The original dataset contains Wikipedia abstracts (context), their titles (title), and a string (triplets) representing multiple knowledge graph relations found in the text.

This version transforms the original dataset by expanding the triplets string so that each row corresponds to a single relation triplet. This format is more convenient for tasks focusing on individual relation facts, such as relation extraction training or knowledge graph population.

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.

Configs

• full — the complete expanded triplet dataset (~4.4M rows), fields: title, text, entity1, entity2, relation.
• subsampled — a smaller subset (~150K rows) used in the paper, with the same fields plus the prompt_0_shot / prompt_2_shot / prompt_5_shot instruction-prompt columns.

from datasets import load_dataset

full = load_dataset("Despina/rebel-dataset", "full")
subsampled = load_dataset("Despina/rebel-dataset", "subsampled")

Data Source and Processing

• Source Dataset: Babelscape/rebel-dataset
• Processing Steps:
  1. The original dataset was loaded using the Hugging Face datasets library.
  2. The triplets feature string in each original row was parsed. The parsing assumed the format <triplet> head <subj> relation <obj> tail [...] based on the example provided in the original REBEL work (note: the exact delimiter tokens might vary slightly, but the head <subj> relation <obj> tail structure was targeted).
  3. For each successfully parsed triplet (head, relation, tail) within a triplets string, a new row was created in the processed dataset.
  4. The original context feature was retained and renamed to text.
  5. The original title feature was retained.
  6. The extracted head was stored as entity1.
  7. The extracted tail was stored as entity2.
  8. The extracted relation name was stored as relation.

Dataset Structure

The dataset contains the standard train, validation, and test splits.

Data Fields

Each row in the dataset has the following fields:

• text (string): The original Wikipedia abstract context.
• title (string): The title of the Wikipedia abstract.
• entity1 (string): The head entity of the relation triplet.
• entity2 (string): The tail entity of the relation triplet.
• relation (string): The name/label of the relation connecting entity1 and entity2.

The subsampled config additionally includes:

• prompt_0_shot (string): Zero-shot instruction prompt (task instructions + the input text).
• prompt_2_shot (string): The same prompt with 2 in-context demonstrations prepended.
• prompt_5_shot (string): 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.

Example Row

{
  "text": "Gustavus Adolphus College is a private liberal arts college in St. Peter, Minnesota. It was founded in 1862 by Swedish Americans led by Eric Norelius and is affiliated with the Evangelical Lutheran Church in America. The college retains affiliations with the Lutheran Church and the Church of Sweden.",
  "title": "Gustavus Adolphus College",
  "entity1": "Gustavus Adolphus College",
  "entity2": "Evangelical Lutheran Church in America",
  "relation": "religion"
}

Usage

You can load the dataset using the datasets library:

from datasets import load_dataset

ds = load_dataset("Despina/rebel-dataset", "subsampled")

# Access splits
train_data = ds['train']
validation_data = ds['validation']
test_data = ds['test']

# Example: Print the first example from the training set
print(train_data[0])

Licensing Information

This dataset is distributed under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0), consistent with the licensing of the original REBEL dataset.

Citation

If you use this dataset, please cite our paper along with the original REBEL 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{huguet-cabot-navigli-2021-rebel,
    title = "{REBEL}: Relation Extraction By End-to-end Language generation",
    author = "Huguet Cabot, Pere Llu{\'\i}s  and
      Navigli, Roberto",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2021",
    month = nov,
    year = "2021",
    address = "Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "[https://aclanthology.org/2021.findings-emnlp.204](https://aclanthology.org/2021.findings-emnlp.204)",
    doi = "10.18653/v1/2021.findings-emnlp.204",
    pages = "2370--2381",
}