Fine-tuning is both over-recommended and under-used. It is over-recommended because a better prompt or better retrieval solves most problems more cheaply. It is under-used because for a narrow, repetitive, well-defined task, a small fine-tuned model can be faster, cheaper and more consistent than any prompt against a large one. This course teaches you to tell the two situations apart, then to do the work properly.
The first module is a decision framework with a genuine escape hatch: several worked examples end with the recommendation not to fine-tune, and the reasoning is spelled out. Once you have a task that qualifies, the real work begins, and most of it is data. You will build a dataset from production examples, clean and deduplicate, and hold out a test set that never touches training.
Training itself uses parameter-efficient methods, principally LoRA, on small open models you can run on a single rented GPU or a decent local card. You will study rank, learning rate and epochs through their effect on loss curves, then overfit a model deliberately so you can recognise the shape of it.
Evaluation is where the course spends its final third. You compare the fine-tune against a strong prompted baseline on held-out data, measure latency and cost per thousand requests, then serve the model behind a stable API with versioning and a rollback path.