Detecting Unreliable News with Fine-Tuned RoBERTa

This project fine-tunes the RoBERTa model for detecting unreliable news articles. The application classifies news articles as reliable or unreliable by leveraging a pre-trained transformer model and a labeled dataset. The backend is implemented using Flask and deployed on Render.com, with the fine-tuned model stored on Hugging Face Space.

Features

• Model: Fine-tuned roberta-base for binary classification.
• Dataset: Labeled dataset with news article attributes: title, text, and reliability label.
• Deployment: Flask-based API hosted on Render; model stored on Hugging Face for easy access.
• Evaluation: Achieves over 99% accuracy and F1 score on the validation dataset.

Table of Contents

1. Installation
2. Usage
3. Project Workflow
4. Model Training
5. Results
6. Deployment
7. Acknowledgments

---

Installation

1. Clone the repository:

   git clone https://github.com/username/unreliable-news-detector.git
   cd unreliable-news-detector

2. Install dependencies:

   pip install -r requirements.txt

3. Set up Hugging Face authentication:

   export HF_TOKEN=your_huggingface_token

4. Configure Flask environment variables:

   export FLASK_APP=app.py
   export FLASK_ENV=development

---

Usage

Run the Flask server:

flask run

Access the API at http://localhost:5000 or the deployed version on Render.

API Endpoints

• POST /predict
  Input: JSON with title and text.
  Output: Predicted label (0 = Reliable, 1 = Unreliable) and confidence score.

---

Project Workflow

1. Data Preprocessing:

   • Cleaned text by removing special characters and URLs.
   • Combined title and text fields for holistic context.
   • Removed duplicates and balanced classes.

2. Data Splitting:

   • 80/20 split for training and validation.
   • Ensured no overlap between sets.

3. Model Training:

   • Used Hugging Face's Trainer API.
   • Fine-tuned RoBERTa for 2 epochs with a learning rate of 2e-5.

4. Evaluation:

   • Metrics: Accuracy and F1 Score.
   • Plotted confusion matrix and performance graphs.

---

Model Training

Hyperparameters:

• Learning Rate: 2e-5
• Batch Size: 16
• Epochs: 2
• Weight Decay: 0.01

Performance:

• Training Loss: 0.0556
• Validation Accuracy: 99.35%
• Validation F1 Score: 99.34%

---

Results

Confusion Matrix:

	Predicted Reliable	Predicted Unreliable
Actual Reliable	2055	11
Actual Unreliable	15	1944

Visualization:

• Plotted training loss and evaluation accuracy over epochs.
• Highlighted the model's strong generalization capabilities.

---

Deployment

Backend:

• Flask REST API serving predictions via POST requests.

Hosting:

• Backend deployed on Render.com.
• Model weights stored on Hugging Face Space for easy accessibility.

---

Acknowledgments

• Hugging Face for providing pre-trained RoBERTa and the Trainer API.
• Render.com for hosting the Flask backend.

---