Your company has millions of product reviews and a product team asking one question: what are customers actually complaining about? Star ratings won’t tell you. Sending every review to a frontier LLM will tell you, but at a cost and latency nobody will approve for a weekly report.

In this hands-on workshop, you’ll build the pattern teams use in production: a big model labels, a small model scales, and SQL answers the question. Starting from a public dataset of Amazon product reviews on the Hugging Face Hub, you’ll use DuckDB to sample and explore the data. Then you’ll have a large LLM tag a few thousand reviews with product aspects (battery, fit, durability, shipping) and sentiment. You’ll LoRA fine-tune a small open LLM to return those labels as JSON, run it over the full dataset, and write the results back to Parquet on the Hub.

Then the fun part: with every review turned into structured columns, you’ll use DuckDB to answer real Voice-of-the-Customer questions. Which brands get the most “arrived broken” complaints? Did battery complaints spike after a product revision? Finally, you’ll ship the pipeline as an interactive app on Hugging Face Spaces.

Along the way, you’ll learn when fine-tuning is the right tool and when it isn’t. Fine-tuning teaches a model a task, not facts. That’s why it’s the right tool for turning messy text into structured data, and the wrong one for helping a model “know” your reviews.

You’ll leave with a working, reusable pipeline and a clear mental model for choosing between prompting, RAG and fine-tuning.