📊 SMS Fraud Detection and Reporting using LLM

🧠 Project Overview

This project implements an end-to-end pipeline for detecting SMS spam using LLM-based embeddings (Mistral), interpretable machine learning, and risk-aware reporting.

It includes:

• Exploratory Data Analysis (EDA)
• Embedding generation using ollama Mistral model
• Random Forest classifier with performance evaluation
• LIME explanations for interpretability
• Executive-level HTML/PDF reporting using LLM-generated narrative

---

🗂️ Repository Structure

├── data/
│   ├── spam.csv                 # Raw dataset (Kaggle UCI SMS Spam)
│   ├── model_metrics.csv        # Saved model evaluation metrics
├── plots/
│   ├── *.png                    # Visuals from EDA, LIME, Confusion Matrix
├── reports/
│   ├── fraud_detection_report.html
│   ├── fraud_detection_report.pdf
├── notebooks/
│   ├── 01_EDA.ipynb
│   ├── 02_llm_finetuning_prediction.ipynb
│   ├── 03_llm_executive_report.ipynb
├── README.md                   

---

📥 Installation

1. Clone the repository:

   git clone https://github.com/sumitdeole/LLM_text_data.git
   cd LLM_text_data

2. Create environment and install dependencies:

   conda create -n sms-fraud python=3.10 -y
   conda activate sms-fraud
   pip install -r requirements.txt

3. Install additional system dependencies:

   • WeasyPrint requires GTK3 runtime for Windows
   • Add C:\Program Files\GTK3-Runtime Win64\bin to your system PATH

---

⚙️ Notebooks

1. EDA + Feature Engineering

• Loads and visualizes data
• Generates word clouds and top spam unigrams

2. Model Training with Mistral

• Generates 4096-D embeddings using ollama's Mistral
• Trains a balanced Random Forest
• Saves metrics and plots

3. Executive Reporting

• Feeds metrics and text features to LLM for narrative
• Renders HTML/PDF executive report with visuals

---

📈 Example Outputs

• Confusion Matrix
• LIME Explanation

---

📄 License

MIT License

---

⭐ Star this repo

If you find this project helpful, feel free to give it a ⭐ on GitHub!