Dataset Card for Sentence-Level Re-DocRED with Full Relation Names

Dataset Description

This dataset is a transformed version of the tonytan48/Re-DocRED dataset, which itself is based on the original DocRED (Document-Level Relation Extraction) dataset.

The key transformations applied are:

1. Sentence-Level Conversion: The original dataset is document-level, meaning each sample contains multiple sentences, entities, and relations within a document. This version unravels the data so that each row corresponds to a single relation instance within one specific evidence sentence. If a relation has multiple evidence sentences listed in the original data, it will appear as multiple rows in this dataset, one for each evidence sentence.
2. Text Cleaning: The text field (containing the sentence) has been processed to normalize spacing around punctuation and specific quote patterns (e.g., Zest Air' ' ) becomes Zest Air''), Airport ' s becomes Airport's).
3. Full Relation Names: The relation feature contains the full textual name of the relation (e.g., "country", "place of birth", "head of government") mapped from the original Wikidata Property IDs (e.g., "P17", "P19", "P6") using a predefined mapping. If a relation ID from the source data was not found in the mapping, the original ID is retained as the value.
4. Instruction Prompts: Each row additionally carries ready-to-use instruction prompts (prompt_0_shot, prompt_2_shot, prompt_5_shot) so the dataset can be used directly for prompt-conditioned fine-tuning and evaluation of (small) language models.

This format is often more suitable for sentence-based relation extraction models.

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.

Source Data:

• Based on: tonytan48/Re-DocRED on Hugging Face Hub.
• Original DocRED: https://github.com/thunlp/DocRED

Dataset Structure

Data Instances

Each instance represents a potential relation between two entities within a specific sentence.

Example Record: (Example data; actual values may vary)

{
  'text': "Zest Airways, Inc. operated as AirAsia Zest (formerly ''Asian Spirit and Zest Air''), was a low-cost airline based at the Ninoy Aquino International Airport's in Pasay City, Metro Manila in the Philippines.",
  'title': "Zest Airways",
  'entity1': "Zest Airways",
  'entity2': "Philippines",
  'entity1Type': "ORG",
  'entity2Type': "LOC",
  'relation': "country" # Mapped from original P17
}

Data Fields

• text (string): The cleaned text of a single sentence potentially containing evidence for the relation.
• title (string): The title of the original Wikipedia document from which the sentence originates.
• entity1 (string): The text/name of the head entity involved in the relation. Note: This is typically derived from the first mention listed in the vertexSet of the source data for the corresponding entity index.
• entity2 (string): The text/name of the tail entity involved in the relation. Note: Derived similarly to entity1.
• entity1Type (string): The semantic type of the head entity (e.g., PERSON, ORG, LOC). Note: Derived similarly to entity1.
• entity2Type (string): The semantic type of the tail entity (e.g., PERSON, ORG, LOC). Note: Derived similarly to entity1.
• relation (string): The full name of the relation between entity1 (head) and entity2 (tail) (e.g., 'country', 'place of birth'). If the original relation ID (like 'PXXX') was not found in the predefined mapping used during creation, the original ID string is used here as a fallback.
• prompt_0_shot (string): Zero-shot instruction prompt (task instructions + the input sentence).
• 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.

Data Splits

The dataset is divided into the same splits as the original Re-DocRED dataset:

• train: Training data.
• validation: Validation data.
• test: Test data.

(Note: Relation labels in the original DocRED test set are typically held out. This dataset includes relations from the test split if they were present in the source tonytan48/Re-DocRED test split; otherwise, the test split might be empty or lack meaningful relation labels depending on the source version).

Dataset Creation

This dataset was generated by processing the tonytan48/Re-DocRED dataset using a Python script with the Hugging Face datasets library. The script iterates through each document, identifies relation labels (labels), and for each relation, iterates through its specified evidence sentences (evidence). For each evidence sentence associated with a relation, a new record is created containing the sentence text, document title, head/tail entity details (name and type, derived from the first mention in the original vertexSet), and the mapped relation name. Text cleaning regex was applied to the sentence text.

Usage

You can load the dataset using the Hugging Face datasets library:

from datasets import load_dataset

repo_name = "Despina/re-docred"
dataset = load_dataset(repo_name)

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

# Example access
print(f"First training example:\n{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 DocRED dataset.

Citation Information

If you use this dataset, please cite our paper along with the original DocRED and Re-DocRED papers:

@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{yao2019docred,
  title={DocRED: A Large-Scale Document-Level Relation Extraction Dataset},
  author={Yao, Yuan and Ye, Deming and Li, Peng and Han, Xu and Lin, Yankai and Liu, Zhenghao and Liu, Zhiyuan and Huang, Lixin and Zhou, Jie and Sun, Maosong},
  booktitle={Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics},
  pages={764--777},
  year={2019},
  publisher={Association for Computational Linguistics}
}

@inproceedings{tan2022revisiting,
  title={Revisiting DocRED-Addressing the False Negative Problem in Relation Extraction},
  author={Tan, Qingyu and Xu, Lu and Bing, Lidong and Ng, Hwee Tou and Aljunied, Sharifah Mahani},
  booktitle={Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing},
  pages={8472--8487},
  year={2022}
}