Fake News Detection using RoBERTa + LoRA Fine-Tuning

A high-performance Fake News Detection pipeline built on RoBERTa with LoRA (Low-Rank Adaptation) fine-tuning. The project demonstrates how parameter-efficient fine-tuning can deliver strong results even on limited compute, with a focus on cross-domain generalization across FakeNews-Kaggle and LIAR datasets.

---

Table of Contents

• Overview
• Project Goals
• Pipeline Overview
• Repository Structure
• Setup Instructions
• Usage
• Results
• Key Insights
• Future Work
• License

---

Overview

This project applies Parameter-Efficient Fine-Tuning (PEFT) using LoRA adapters on RoBERTa-base, enabling efficient adaptation of large language models for misinformation detection. It evaluates both in-domain performance (FakeNews-Kaggle) and out-of-domain generalization (LIAR dataset) to measure model robustness.

---

Project Goals

• Develop an end-to-end reproducible pipeline for fake news detection.
• Fine-tune RoBERTa efficiently using LoRA adapters.
• Benchmark cross-domain performance using both FakeNews and LIAR datasets.
• Maintain lightweight training with minimal GPU memory footprint.
• Provide structured evaluation metrics and results visualization.

---

Pipeline Overview

Stage	Description
1. Data Fetching	Downloads and preprocesses FakeNews-Kaggle and LIAR datasets.
2. Preprocessing	Cleans text, encodes labels, and prepares train/val/test splits.
3. Model Setup	Initializes RoBERTa-base with LoRA adapters using the PEFT library.
4. Training	Fine-tunes model using mixed precision and checkpoint saving.
5. Evaluation	Computes metrics (Accuracy, Precision, Recall, F1) across domains.

---

Repository Structure

Misinformation-Classifier/
├── src/
│   ├── data/
│   │   ├── fetch_datasets.py       # Fetch and prepare datasets (FakeNews, LIAR)
│   │   ├── datasets.py             # PyTorch Dataset class
│   │   └── preprocess.py           # Cleaning, tokenization, and encoding
│   │
│   ├── models/
│   │   ├── text_model.py           # RoBERTa + LoRA architecture
│   │   └── train.py                # Training loop and checkpoint saving
│   │
│   ├── evaluate.py                 # Evaluation script for performance metrics
│   └── utils/
│       ├── metrics.py              # Metric computation utilities
│       └── seed.py                 # Reproducibility and seed setup
│
├── notebooks/
│   ├── 01_data_exploration.ipynb   # Dataset analysis and visualization
│   ├── 02_model_training.ipynb     # LoRA fine-tuning workflow
│   └── 03_model_evaluation.ipynb   # Performance evaluation and confusion matrix
│
├── data/
│   ├── raw/                        # Raw data (downloaded automatically)
│   └── processed/                  # Cleaned and tokenized data splits
│
├── models/
│   ├── roberta_lora_multidomain_best/   # Best-performing model checkpoint
│   ├── roberta_lora_multidomain_steps/  # Step-based checkpoints
│   ├── roberta_lora_multidomain_merged/ # Final merged adapter model
│   ├── roberta_lora_multidomain_history.csv # Training history log
│   └── roberta_lora_multidomain_evaluation_summary.csv # Evaluation summary
│
├── requirements.txt                # Python dependencies
├── Makefile                        # Reproducible training commands
└── README.md                       # Project documentation

---

Setup Instructions

1. Clone the Repository

git clone https://github.com/ashbix23/Misinformation-Classifier.git
cd Misinformation-Classifier

2. Create a Virtual Environment

conda create -n fakenews python=3.10 -y
conda activate fakenews

3. Install Dependencies

pip install -r requirements.txt

4. Download and Preprocess Data

python -m src.data.fetch_datasets

This script downloads, cleans, and processes both FakeNews-Kaggle and LIAR datasets, saving ready-to-use files in data/processed/.

---

Usage

1. Explore the Data

jupyter notebook notebooks/01_data_exploration.ipynb

2. Train the Model

python -m src.train --model roberta-base --lora --datasets FakeNews LIAR

3. Evaluate the Model

jupyter notebook notebooks/03_model_evaluation.ipynb

This notebook generates accuracy, precision, recall, F1, and confusion matrices for each dataset.

---

Results

Dataset	Accuracy	Precision	Recall	F1
FakeNews	0.773	0.846	0.773	0.763
LIAR	0.443	0.754	0.443	0.273

Interpretation

• Excellent in-domain performance on FakeNews-Kaggle.
• Significant drop in cross-domain generalization for LIAR (short factual claims).
• Model learns strong contextual and linguistic cues for misinformation.

---

Key Insights

• LoRA fine-tuning reduces memory and compute cost while preserving model quality.
• Cross-domain adaptation remains challenging due to dataset structural differences.
• Clean preprocessing and normalization dramatically improve results.
• Mixed Precision Training (AMP) enhances training efficiency on GPUs.

---

Future Work

• Domain-adaptive fine-tuning for short-form factual claims.
• Add more datasets for multi-domain robustness.
• Build an inference API using FastAPI or Gradio for quick model testing.

---

License

This project is licensed under the MIT License. You are free to use, modify, and distribute it for academic or research purposes.

---

Acknowledgments

• Hugging Face Transformers
• PEFT (Parameter-Efficient Fine-Tuning)
• PyTorch
• FakeNews-Kaggle Dataset
• LIAR Dataset