LLM-301

Fine-tuning Open Models

When and how to fine-tune an open model, judged against the alternative every senior engineer should weigh: not doing it.

ABOUT THIS COURSE

What you will learn.

Fine-tuning is one of the easiest mistakes to make in applied AI. It looks like progress, it produces metrics, and it generates an artifact you can point at — and in roughly seventy percent of the cases we see, it should not have been the first choice. This course is built around the discipline of asking, honestly, whether fine-tuning is the right tool, and then doing it well when the answer is yes.

Over six weeks, we cover the modern open-model landscape and what is realistic to expect from each family, the supervised fine-tuning recipe end to end, parameter-efficient methods (LoRA, QLoRA, DoRA), evaluation under distribution shift, deployment cost vs. base-model cost, and the operational disciplines required to keep a fine-tuned model honest. You will fine-tune one model from scratch as a working exercise and then revisit that decision with the eyes of someone who has now done it.

This course is the technical companion to the engineering judgment we teach in AI-201. It is also the only Solutech course that involves a GPU.

WHAT YOU’LL BUILD

Four substantial projects.

Project 01

A “should we fine-tune?” decision memo

Write a defensible memo for a real use case, then defend it against a cohort review.

Project 02

A LoRA fine-tune end to end

Take a base model, prepare data, train a LoRA adapter, evaluate honestly, and deploy.

Project 03

A pre/post evaluation study

Construct an eval that distinguishes real improvement from memorization.

Project 04

A cost-and-latency comparison

Compare the fine-tuned model against the base model plus prompting on cost, latency, and quality.

CURRICULUM

Week by week.

FIT

Who this is for — and who it is not.

For you if

Engineers who have been asked to “fine-tune our model” and want to do it well.
Engineers who want the technical depth to push back on a fine-tuning project that shouldn’t happen.
ML engineers from non-LLM backgrounds entering the LLM stack.

Probably not for you if

Pre-training researchers — this is applied, not foundational.
Engineers without GPU access; arrangements can be made but expect setup work.
People looking to fine-tune frontier closed models — most providers do not allow what this course teaches.

YOUR INSTRUCTOR

Taught by an operator.

ML Systems Engineer

Tomasz Kowalski

Tomasz spent the first ten years of his career writing compilers — first at Intel, then at a chip startup that no longer exists — before pivoting into ML systems. He maintains an inference framework used by a handful of well-known products and has strong opinions about KV-cache layouts. His course is the only one in the Solutech catalog with a homework set involving CUDA.

FAQ

Questions we’re asked often.

LLM-301 · Next cohort starts soon