🧠 DBpediaOntoTrain: A Quality-Segmented Ontology Dataset for LLM Pretraining

📘 Overview

DBpediaOntoTrain is a dataset of 1,766 OWL ontologies in Turtle format, extracted from DBpedia Archivo and prepared for continual pretraining of Large Language Models (LLMs) in ontology generation and completion tasks.

Each ontology is analyzed using a set of semantic quality metrics, tokenized using the LLaMA 3.2 tokenizer, and sorted by Quality Score (QS). The dataset includes cumulative token counts and percentages, allowing precise and reproducible slicing for quality-aware training.

---

📦 Dataset Contents

• data.json: A JSON file where each entry contains:
  • File Name: name of the ontology file (.ttl)
  • plain_text: raw ontology content in Turtle syntax
  • PD: Property Density by Class
  • NTR: Non-Taxonomic Relations per Class
  • SC: Subclasses per Class
  • PD_norm, NTR_norm, SC_norm: min-max normalized versions of the above metrics
  • QS: Quality Score (PD_norm + NTR_norm + SC_norm)
  • Token Count: number of tokens computed using the LLaMA 3.2 tokenizer
  • Token Count Accumulation: cumulative token count (sorted by descending QS)
  • Percentage of Token Count Accumulation: running percentage of total tokens across all ontologies

The dataset is sorted in descending order by Quality Score (QS), enabling easy extraction of quality-based subsets (e.g., Q1, Q1,2, etc.).

---

⚠️ Loading the Dataset

The standard datasets.load_dataset() function from the Hugging Face datasets library does not work with this dataset, likely due to format or hosting issues.

However, you can easily load it using Python's built-in json module:

import json

with open('path/to/data.json', 'r', encoding='utf-8') as f:
    data = json.load(f)

This will give you a list of dictionary entries, each representing one ontology and its associated quality metrics, ready for filtering or slicing based on your training needs.

---

📊 Quality Metrics

Each ontology is scored with:

Metric	Description
PD	Property Density — properties per class
NTR	Non-Taxonomic Relations — domain-specific relations per class
SC	Subclass Count — hierarchical depth
QS	Sum of normalized PD, NTR, SC

These metrics reflect semantic modeling richness rather than raw size.

---

🧪 Intended Use

• Continual pretraining of LLMs on semantic data
• Research in ontology learning, alignment, enrichment
• Studying the effect of data quality on model generalization and reasoning

This dataset supports the research study:

Enhancing LLM Ontology Generation: The Role of Quality Semantic Data
Miquel Canal-Esteve, Yoan Gutiérrez, José Abreu-Salas (submitted to ICT Express, 2025)

---

🛠️ Tokenization

• Tokenized using LLaMA 3.2-1B tokenizer
• Total tokens: 1.25 billion
• Cumulative token fields allow extracting top-N% token subsets based on QS
• Token overlap and LLM input chunking are described in the accompanying paper

---

💡 Reproducibility

The repository includes:

• Metric calculation scripts using rdflib
• Tokenization scripts with Hugging Face libraries
• Pretraining configs and logs

Repository:
👉 https://github.com/miquelcanalesteve/LLM4Onto/

---

📄 Citation

@misc{canal2025dbpediaontotrain,
  author    = {Miquel Canal-Esteve and Yoan Gutiérrez and José Abreu-Salas},
  title     = {DBpediaOntoTrain: A Quality-Segmented Ontology Dataset for LLM Pretraining},
  year      = {2025},
  url       = {https://github.com/miquelcanalesteve/LLM4Onto/}
}