YAGO 4.5 Dataset (English subset for LLM fine-tuning)

Dataset Description

This datasets contains triples filtered from yago-facts.ttl and yago-beyond-wikipedia.ttl in the YAGO 4.5 dataset. The SPARQL query used to filter the triples is in raw/filter.sparql. This represents a subset of the YAGO 4.5 dataset maintaining only English labels.

I remapped some schema.org URIs to use the http://yago-knowledge.org/resource/ which were not present in the schema.org vocabulary. I also removed schema:sameAs and owl:sameAs relations from this dataset, as well as triples with xsd:anyURI object literals, as my goal is to use this dataset for fine-tuning a large language model for knowledge graph completion and I do not want to train the base model to predict these kind of relations.

Overview

YAGO 4.5 is the latest version of the YAGO knowledge base. It is based on Wikidata — the largest public general-purpose knowledge base. YAGO refines the data as follows:

All entity identifiers and property identifiers are human-readable.
The top-level classes come from schema.org — a standard repertoire of classes and properties maintained by Google and others. The lower level classes are a careful selection of the Wikidata taxonomy.
The properties come from schema.org.
YAGO 4.5 contains semantic constraints in the form of SHACL. These constraints keep the data clean, and allow for logical reasoning on YAGO.

Dataset Structure

The dataset is structured as follows:

raw/yago-taxonomy.ttl: Contains the rdfs:subClassOf relations for YAGO and the prefix mappings for the N-Triples.
raw/facts.tar.gz: Compressed file containing chunks of the dataset in N-Triples format, representing the factual knowledge in YAGO.

Features

Each RDF triple in the dataset is represented with the following features:

subject: The subject of the triple, representing the entity.
predicate: The predicate of the triple, representing the relationship between the subject and object.
object: The object of the triple, representing the entity or value linked by the predicate.

Splits

The dataset is logically divided into multiple chunks, each containing a subset of RDF triples. Users can load specific chunks or the entire dataset based on their requirements.

Usage

Loading the Dataset

The dataset can be loaded using the Hugging Face datasets library as follows:

from datasets import load_dataset

dataset = load_dataset('wikipunk/yago-4.5-en', num_proc=4)

Accessing the YAGO Taxonomy File

The yago-taxonomy.ttl file can be accessed and loaded in every process as follows:

from rdflib import Graph

taxonomy_graph = Graph()
taxonomy_graph.parse('raw/yago-taxonomy.ttl', format='turtle')

Additional Information

Licensing

The YAGO 4.5 dataset is available under the Creative Commons Attribution-ShareAlike 3.0 license.

Citation

If you use the YAGO 4.5 dataset in your work, please cite the following publication:

@article{suchanek2023integrating,
  title={Integrating the Wikidata Taxonomy into YAGO},
  author={Suchanek, Fabian M and Alam, Mehwish and Bonald, Thomas and Paris, Pierre-Henri and Soria, Jules},
  journal={arXiv preprint arXiv:2308.11884},
  year={2023}
}