Module 3 · Project 1: Banking Customer Query Classifier (fine-tuning BERT)

OBJECTIVE: build an app that reads a customer's message ('my card hasn't arrived yet') and routes it to the right one of 77 banking intents — or to a human agent when unsure — and learn text classification and fine-tuning while building it. WHAT WE COVER (in project order): • Defining the problem and the score • Exploring the data and labelling your own text examples • Preparing text for a Transformer: splits, label IDs, tokenisation • A fast classic baseline (TF-IDF + logistic regression) • Fine-tuning DistilBERT with the Hugging Face Trainer • Evaluation: accuracy, macro-F1, per-intent results, confidence threshold, error analysis • Deployment: model on the Hugging Face Hub + Gradio app • Optional next step: entity extraction (NER) with the same tools AFTER THIS MODULE YOU CAN: • Label and prepare any text classification dataset • Fine-tune a Transformer and prove it beats a simple baseline • Ship a text classifier with a safe 'hand over to a human' rule DATASET: Banking77 (13,083 real customer messages, 77 intents) + 100 messages you write and label yourself. TOOL: Hugging Face Transformers.