|
|
--- |
|
|
title: Headlne |
|
|
emoji: 🔥 |
|
|
colorFrom: indigo |
|
|
colorTo: pink |
|
|
sdk: gradio |
|
|
sdk_version: 5.23.1 |
|
|
app_file: app.py |
|
|
pinned: false |
|
|
--- |
|
|
|
|
|
Bias Bin: Bias Detection and Mitigation in Language Models |
|
|
|
|
|
Bias Bin is an interactive Gradio-based web application for detecting and mitigating gender bias in narrative text. It uses a fine-tuned BERT model and counterfactual data augmentation techniques to highlight and analyze bias in NLP outputs. |
|
|
|
|
|
🧠 Project Overview |
|
|
|
|
|
This tool allows users to: |
|
|
• Detect gender bias in input text using a BERT-based classification model. |
|
|
• Explore counterfactual predictions by swapping gendered terms. |
|
|
• Visualize bias scores to understand model behavior. |
|
|
• Demonstrate bias mitigation through gender-swapped text examples. |
|
|
|
|
|
This project was developed as part of a university coursework in Deep Learning & Generative AI. |
|
|
|
|
|
📁 Repository Contents |
|
|
• app.py – Main Python file to launch the Gradio web app. |
|
|
• Evaluation&Results.ipynb – Notebook with experiments, model evaluations, and visualizations. |
|
|
• fine_tuned_model.zip – Zip file containing the fine-tuned BERT model (must be extracted). |
|
|
• requirements.txt – List of Python dependencies. |
|
|
|
|
|
⚙️ Setup Instructions |
|
|
1. Clone the Repository |
|
|
|
|
|
git clone https://huggingface.co/spaces/aryn25/bias.bin |
|
|
cd bias.bin |
|
|
|
|
|
2. Install Dependencies |
|
|
|
|
|
pip install -r requirements.txt |
|
|
|
|
|
3. Extract the Model |
|
|
Unzip the fine_tuned_model.zip file and place the extracted folder in the project root. |
|
|
4. Run the App |
|
|
|
|
|
python app.py |
|
|
|
|
|
5. Open in Browser |
|
|
Visit the Gradio URL printed in the terminal |
|
|
|
|
|
📊 Methodology |
|
|
• Model: Fine-tuned BERT classifier trained on gender-labeled narrative datasets. |
|
|
• Bias Detection: Uses counterfactual data augmentation by swapping gendered words (e.g., “he” → “she”). |
|
|
• Metrics: Bias scores are computed based on prediction discrepancies between original and counterfactual samples. |
|
|
|
|
|
📚 References |
|
|
|
|
|
This project is built using foundational and peer-reviewed research on: |
|
|
• BERT and Transformer models |
|
|
• Gender bias in NLP |
|
|
• Counterfactual data augmentation |
|
|
• Bias mitigation techniques |
|
|
|
|
|
Full citation list available in the project report. |
|
|
|
|
|
📌 Authors |
|
|
|
|
|
Created by Aryan N. Salge and team as part of the Deep Learning & Generative AI coursework at the National College of Ireland. |
|
|
|
|
|
📄 License |
|
|
|
|
|
This project is for educational and research purposes. Please cite appropriately if you use or adapt the work. |
|
|
|
