Galerasnet/pre-conceptual-schemas-alpaca

Pre-Conceptual Schemas Dataset in Alpaca Format

This repository contains the dataset used for fine-tuning the Phi-3.5-PCS-Finetuned model, designed for the interpretation of Pre-Conceptual Schemas (PCS).

The dataset is a result of the master's thesis project "Natural Language Processing in Pre-conceptual Schemas for Representing Knowledge by Using Large Language Models" from the Master's Degree in Systems and Computing Engineering at the University of Nariño.

• —Dataset Curators: Felipe Roa Narváez, Jesús Insuasti, Ph.D.
• —Affiliation: University of Nariño
• —Language: English
• —Dataset Size: 538 examples 

Dataset Summary

This is a dataset for the supervised fine-tuning of language models. Its purpose is to "teach" a model to transform a structured representation of a Pre-Conceptual Schema (a list of semantic triads in JSON format) into a coherent and detailed natural language description.

The dataset was specifically created to address the niche task of interpreting knowledge diagrams that, while useful for humans, pose challenges for automatic computational interpretation.

Supported Tasks

• —text-generation: The dataset is designed for conditional text generation tasks, where the model must generate a descriptive paragraph based on a structured input.

Dataset Structure

Data Instances

The dataset follows the Alpaca format, with each entry being a JSON object containing three fields: 

instruction, input, and output. An example is shown below:

JSON

{

    "instruction": "Interpret the following pre-conceptual schema",
    "input": "[{\"subject\": \"CUSTOMER\", \"predicate\": \"BUYS\", \"object\": \"PRODUCT\"}, {\"subject\": \"PRODUCT\", \"predicate\": \"HAS\", \"object\": \"PRICE\"}, {\"subject\": \"STORE\", \"predicate\": \"SELLS\", \"object\": \"PRODUCT\"}, {\"subject\": \"CUSTOMER\", \"predicate\": \"PAYS\", \"object\": \"STORE\"}, {\"subject\": \"STORE\", \"predicate\": \"HAS\", \"object\": \"INVENTORY\"}]",
    "output": "The diagram illustrates the process by which a customer assesses student performance and provides feedback."

}

(Source: Based on the format described in the research )

Data Fields

• —instruction: (string) A clear and fixed directive that tells the model the task to perform.
• —input: (string) A JSON-formatted string representation of the semantic triads and key relationships extracted from the original pre-conceptual schema.
• —output: (string) The "gold-standard" reference description in natural language. This was generated by a human expert and serves as the ideal response the model should learn to produce.

Data Splits

The consolidated dataset of 538 examples was carefully split for training and evaluation:

• —train: 484 examples (90% of the total), used for the fine-tuning process.
• —test: 54 examples (10% of the total), used to evaluate the model's performance on unseen data.

Dataset Creation

Curation Rationale

The creation of this dataset was motivated by the need for a specific linguistic corpus for the semantic interpretation of pre-conceptual schemas. The goal was to build a high-quality dataset to train a model capable of grasping the relevant semantic and contextual nuances required for this task.

Source Data

The source data for the corpus was primarily pre-conceptual schemas extracted from various academic sources, including master's and doctoral theses from the digital repository of the University of Nariño. This diversity of sources ensures that the dataset encompasses a wide range of structures and complexities from different knowledge domains.

Annotations

The annotation process was a critical step. For each pre-conceptual schema, a human expert crafted a detailed "gold-standard" natural language description. These descriptions were meticulously written to ensure they were complete, coherent, and accurate, capturing the implicit meaning of the schema. These human annotations serve as the quality reference (

output) for the model's training.

Citation

If you use this dataset in your research, please cite the original work:

@mastersthesis{RoaInsuasti2025,
  author  = {Felipe Roa Narváez, Jesús Insuasti},
  title   = {Natural Language Processing in Pre-conceptual Schemas for Representing Knowledge by Using Large Language Models},
  school  = {Universidad de Nariño},
  year    = {2025},
  address = {San Juan de Pasto, Colombia}
}