kjacone/swahili_cake_boss

Show how to fine-tune Gemma for handling specific email inquiries a Swahili bakery business might receive. # Gemma Fine-Tuning for Swahili Bakery Email Management This project demonstrates how to fine-tune Gemma to handle customer email requests for a Swahili bakery business. The model will process and respond to inquiries about orders, pricing, and bakery services in Swahili. ## Overview The system handles common customer interactions including: - Product availability and pricing inquiries - Custom cake orders - Delivery scheduling - Special event catering requests - Payment and booking confirmations ## Prerequisites - Python 3.9+ - CUDA-compatible GPU (recommended) - Access to Gemma API or model weights - Basic understanding of machine learning and NLP ## Installation 1. Set up your Python environment: ```bash python -m venv venv source venv/bin/activate # On Windows: venv\Scripts\activate ``` 2. Install required packages: ```bash pip install torch transformers datasets pandas numpy tqdm ``` ## Dataset Preparation Create a training dataset with paired customer emails and ideal responses: 1. Structure your data as a CSV file with these columns: - `email_text`: Customer inquiry in Swahili - `response_text`: Appropriate response - `intent`: Email classification (optional) - `entities`: Key information extracted (optional) Example dataset entry: ```csv email_text,response_text,intent "Habari, ningependa kuagiza keki ya birthday. Bei gani?","Karibu! Keki zetu za birthday zinaanza bei ya TSh 30,000. Tafadhali tujulishe ukubwa na mapambo unayopenda ili tukupatie bei kamili.",order_inquiry "Je, mna home delivery?","Ndio, tunatoa huduma ya home delivery ndani ya Dar es Salaam. Bei ya delivery ni kuanzia TSh 5,000 kutegemea umbali.",service_inquiry ``` 2. Save as `swahili_bakery_dataset.csv` ## Model Fine-Tuning ```python from transformers import ( AutoModelForSeq2SeqLM, AutoTokenizer, Seq2SeqTrainingArguments, Seq2SeqTrainer, DataCollatorForSeq2Seq ) from datasets import load_dataset # Load model and tokenizer model_name = "google/gemma-7b" tokenizer = AutoToke …