hafsa-yahyaa/Kenyan-Bank-Customer-Support-AI

Fine-tuning BERT for Swahili/Sheng Customer Support # 🇰🇪 Localized AI: Kenyan Customer Support Classifier A Natural Language Processing (NLP) project that fine-tunes a Large Language Model (DistilBERT) to understand Kenyan code-mixed text (Swahili, English, and Sheng). ## Project Overview Global AI models often struggle with local context. If a user says *"Luku imeisha"* (Power tokens are finished), a standard US-trained model might confuse it for a general statement. This project solves that by fine-tuning **DistilBERT** on a custom dataset of Kenyan customer support queries to classify them into actionable intents. ## Objectives 1. **Create a Local Dataset:** Curated 100+ samples mixed with Swahili/Sheng. 2. **Fine-Tune an LLM:** retrained `distilbert-base-uncased` for classification. 3. **Demonstrate Local Adaptation:** Prove the model understands terms like *M-Pesa, Sambaza, and Bundles*. ## Project Structure * `fine_tune_bert.ipynb`: The Google Colab notebook containing the training pipeline. * `customer_support_data.csv`: The custom dataset with 4 categories of intent. * `train_model.py`: Python script version of the training loop. ## Categories & Performance The model classifies text into 4 categories: * **Billing Issue** (e.g., "Tokens zangu hazija reflect") * **Transaction Reversal** (e.g., "Wrong number, reverse please") * **Network Issue** (e.g., "Netiko iko down") * **General Inquiry** (e.g., "Mnafungua saa ngapi?") **Results:** * **Accuracy:** ~65% (on a small dataset of 100 samples) * **Baseline Comparison:** significantly outperforms random guessing (25%) ## How to Run 1. Open `fine_tune_bert.ipynb` in Google Colab. 2. Upload `customer_support_data.csv` to the session storage. 3. Run all cells to fine-tune the model. ## Author **Hafsa Yahya Mohamud 169440** *Bachelor of Science in Statistics and Data Science*