Kachielite/mpesa-sms-llm-training-pipeline

# MPESA SMS LLM Training Pipeline A comprehensive end-to-end pipeline for processing MPESA SMS transaction data and fine-tuning Large Language Models (LLMs). This project transforms raw SMS messages into structured training datasets and provides complete tools for fine-tuning models to understand and extract transaction information. ## 🎯 Project Overview This project provides a complete machine learning pipeline for MPESA SMS transaction analysis. From raw SMS data to a trained LLM capable of extracting structured transaction information, this toolkit handles the entire workflow including data preprocessing, anonymization, formatting, and model fine-tuning. ### Key Features - **Data Collection**: Load SMS messages from XML backups - **Privacy Protection**: Comprehensive anonymization of personal information - **Intelligent Parsing**: Extract key transaction fields using regex and NLP techniques - **Flexible Formatting**: Support for both basic and instruct/chat model training formats - **Cloud Integration**: Direct upload to Hugging Face Hub for dataset sharing - **LLM Fine-tuning**: Complete pipeline for training models on MPESA transaction data ✅ - **Model Evaluation**: Tools for assessing model performance on transaction extraction tasks - **Mac M1 Optimized**: Specialized configuration for Apple Silicon training ## 📊 Extracted Fields The system extracts the following key information from each SMS: - `transaction_id` - Unique transaction identifier - `amount` - Transaction amount in KSH - `transaction_type` - Type of transaction (sent, received, withdrawn, airtime, etc.) - `counterparty` - Other party involved in the transaction - `date_time` - Transaction timestamp - `balance` - Account balance after transaction ## 🚀 Getting Started ### Prerequisites - Python 3.8+ - Virtual environment (recommended) - For training: Mac M1 with 16GB RAM (recommended) or similar hardware ### Installation 1. Clone the repository: ```bash git clone github.com …