Fine-tuning Large Language Models for the Extraction of Ecological Relationships Involving Microorganisms and Ants

This repository contains the code, data, and documentation for the Undergraduate Research Project “Fine-tuning large language models for the extraction of ecological relationships involving microorganisms and ants” (FAPESP 2026/12637-3).

The project aims to evaluate and fine-tune large language models (LLMs) for the automatic extraction of ecological relationships between ants (Formicidae) and microorganisms from scientific literature, building a reproducible pipeline that spans from the curation of raw data (GloBI, BioC-PMC) to the construction of a knowledge graph of the identified interactions.

Team

• Undergraduate Research Fellow: Lorena Silva (Biological Sciences, UNESP Rio Claro)
• Advisor: Prof. Dr. Renato Augusto Corrêa dos Santos (CEIS/IB/UNESP Rio Claro)
• External Collaborator: Prof. Dr. Ruben Interian (Institute of Computing/UNICAMP)

Methodology

The project follows a three-phase methodology:

1. Baseline Evaluation — selection and screening of interaction data (GloBI) and evaluation of candidate LLMs based on eligibility criteria (license, availability of model weights, support for LoRA/QLoRA, computational feasibility, active maintenance, and technical documentation).
2. Fine-tuning (LoRA/QLoRA) — supervised fine-tuning of the selected model(s) for relationship extraction from narrative text (BioC-PMC corpus), using annotations based on the SRO (Subject–Relation–Object) scheme.
3. Generalization Testing and Knowledge Graph Construction — evaluation of the fine-tuned model on unseen data and organization of the extracted relationships into a knowledge graph.

Requirements

• Python 3.12.3 with the libraries listed in each requirements.txt .

Installation

Clone this repository:

git clone https://github.com/Lab-LBMM/ic_lorena_silva.git
cd ic_lorena_silva

Repository Structure

scripts/
  process_interactions.py   # main data curation and processing pipeline
cic/                         # step-by-step analysis documentation exported from Notion,
                             # corresponding to the work presented at the CIC
  poster/                    # poster presented at the CIC (PDF)

Documentation

Check out my Wiki page

Acknowledgments

This work was carried out with support from the São Paulo Research Foundation (FAPESP), Brazil (grant numbers 2024/19418-0 and 2026/12637-3).

Questions or Contributions?

For suggestions, bug reports, or collaboration, feel free to open an issue.