🧠 Health and COVID-19 Misinformation Detection – Capstone Project (2025)

Welcome to the official documentation for the AI Capstone Project focused on detecting health-related misinformation. This project leverages Flan-T5 (Large), a state-of-the-art transformer model, fine-tuned on two datasets:

• Health Fact: A general-purpose health misinformation dataset.
• COVID-19 Fake News: A dataset targeting fake news related to COVID-19.

The repository contains code, documentation, datasets (preprocessed), benchmarking comparisons, and a functional user interface (UI) chatbot for classifying input prompts as:

• ✅ True
• ❌ False
• ⚠️ Misleading

---

🧪 1. Objective

This project addresses the growing need to combat health misinformation, with two focal datasets:

• Health Fact Dataset – Used to train a general model for health-related claims.
• COVID-19 Fake News Dataset – Tailored to pandemic-specific misinformation.

I fine-tune the Flan-T5 Large model using full and sub-sample versions of both datasets and compare results with:

• 100-sample, 1,000-sample, and full dataset fine-tuning
• Prompt engineering techniques
• Quantized models for memory optimization

---

🧪 Dataset 1: HealthFact

📂 HealthFact/Health fact Cleaned Datasets/

Contains the following:

• Cleaned_healthfact_traindata.json: Main training data
• cleaned_healthfact_dev.json: Evaluation data
• cleaned_healthfact_test.json: Validation data
• All data is preprocessed, cleaned, and labeled with ["True", "False", "Misleading"]

There is also the reduced datasets:

• 'Reduced_HealthFact_Train.json'
• 'Reduced_HealthFact_Dev.json'
• 'Reduced_HealthFact_Test.json'

📔 Health Fact Fake News Classification and Misinformation Detection.ipynb

This notebook:

• Loads and explores the HealthFact dataset
• Preprocesses data using tokenization, truncation, and padding
• Fine-tunes the Flan-T5 Large model with different configurations:
  • Full dataset
  • 100 samples
  • 1,000 samples
• Includes prompt-engineering templates like:

  "Claim: [text]. Is this true, false, or misleading?"

• Evaluates models using accuracy, precision, recall, and F1-score

📁 HealthFact/finetuned_flan_t5_model/

• Saved fine-tuned HuggingFace model
• Includes:
  • pytorch_model.bin
  • config.json
  • tokenizer_config.json
  • Compatible with HuggingFace’s from_pretrained() loading

---

🧪 Dataset 2: COVID-19 Fake News

📂 Covid19 DataSet/data/

Contains:

• Reduced_Covid_Train.json, Reduced_Covid_Dev.json, Reduced_Covid_Test.json
• Cleaned and pre-labeled COVID-19 specific misinformation
• Labels: True, False

📔 Fake News Classification and Misinformation Detection for COVID-19.ipynb

This notebook:

• Processes the COVID-19 dataset
• Fine-tunes Flan-T5 (Large) with varying sample sizes
• Applies prompt engineering to assess few-shot performance
• Contains evaluation benchmarks to compare:
  • Full dataset training
  • 100-sample and 1,000-sample training
  • Prompt-based classification
• Also includes integration instructions for chatbot deployment

📁 Covid19 DataSet/fine-tuned-flan-t5-covid/

• Output directory for the fine-tuned COVID model
• Contains:
  • Model weights and tokenizer files
  • Usable in chatbot or evaluation pipeline

---

🧑‍💻 Author Yu-Cheng Joshua Lin Deakin Capstone Unit, 2025 Deakin University Email: yjlin@deakin.edu.au GitHub: Jo5hylinn