Module 5 · Project 3: Customer Chat Summariser (fine-tuning T5 for text generation)

OBJECTIVE: build an app that turns a customer–agent chat into a short summary of 1–3 sentences, and learn how Transformers generate text and how to judge generated text while building it. WHAT WE COVER (in project order): • Defining the problem: what a good summary is • The dataset and writing reference summaries and a rating guide • Preparing input–output pairs for an encoder-decoder model • How text is generated: greedy, beam search, sampling (temperature, top-k, top-p) • Baselines: simple rule and zero-shot model • Fine-tuning FLAN-T5 with the Seq2Seq Trainer • Evaluation: ROUGE scores + human rating + checking for made-up facts • Deployment: Gradio summariser app AFTER THIS MODULE YOU CAN: • Fine-tune a Transformer that generates text • Control generation with decoding settings • Evaluate generated text with both metrics and human checks DATASET: SAMSum (about 16,000 chat dialogues with human-written summaries; DialogSum is an alternative) + 20 chats you summarise yourself. TOOL: Hugging Face Transformers.