Send it to fellowship

MIS 752 · Lab 11 Lite · Fine-tuning and LoRA · Book Ch. 18.13 and Ch. 23 · no coding · every patient question here is invented

In medicine

A doctor trained broadly

➜

An instructor teaching students on a training mannequin

a fellowship: cases, with a mentor

➜

A nursing student practicing specialized care

comes back as a specialist

In AI

🧠a general model

➜

📚train it on 100 example answers

➜

🩺answers in your nurse's voice

Fine-tuning teaches a model to sound like your expert. It does not teach it to be your expert, and whoever wrote the examples decides who it serves. This lab shows all three, with a real fine-tune.

👓 LoRA is glasses, not eye surgery. The model underneath stays frozen. A small add-on, here 0.44% of the weights, sits in front of it and changes how it sees your task. Take the glasses off and the model is exactly what it was. You can keep a drawer of them: one for discharge notes, one for billing letters.

📖 Read Ch. 18.13, Prompting, Retrieval, or Fine-Tuning (p. 431) ➜ 📖 Read Ch. 23, Algorithmic Bias and Equity (p. 545) ➜

Same book on Canvas: course files

Photos: Josh Hawkins, Benjamin Richards, Becca Schwartz and Aaron Mayes, UNLV Photo Services, and UNLV Special Collections & Archives. Real UNLV spaces; every patient question in this lab is invented.

Why this matters

The one idea. Fine-tuning trains an existing model on your own examples. It changes how a model talks far faster than what it knows: 100 answers in a discharge nurse's voice teach it her voice in seconds, not her judgment. And whoever wrote the examples decides who it serves.

Why it matters. A model that only sounds better can ship as if it got better, and patients follow a calm, nurse-shaped answer even when it skips the step that matters. For a business, knowing this decides whether training is worth paying for.

By the end you will be able to:

tell a problem of form (voice, format, habits) from one of knowledge, and pick the fix to match;
grade a fine-tune's voice and its advice separately, instead of trusting a demo;
check whose examples a model learned from, and who was left out.

0 · Connect a model

Only section 2 calls a live model. Paste the free OpenRouter key from Lab 1. It stays in this browser tab only: it is never saved and never sent anywhere except OpenRouter.

🔑 OpenRouter API key

Model (if one is busy, pick another)

Show paid models too

1 · Three ways to change how a model behaves

Try them in this order. Each one fixes a different problem.

💬

Prompting

tell it what you want, in words

Always try this first.

free · instant

📄

Retrieval (Lab 8)

hand it documents to read first

It needs facts it does not have, or facts that change.

cheap · no training

🎓

Fine-tuning (today)

train it on your examples

It needs a voice, format or habit it does not have.

a GPU · a casebook of examples

Retrieval for what it KNOWS. Fine-tuning for how it BEHAVES.

🔎 What this shows. The three tools fix different problems, so the first question is never "should we fine-tune?" but "is this a problem of form or of knowledge?" That matters because training is the slowest and most expensive fix, and a team that reaches for it first can spend weeks on something a better prompt or a document lookup would have solved.

2 · Teach it by examples, live

This is what fine-tuning is trying to do, done the quick way: show the model a few example answers with every question. Fine-tuning bakes the same kind of examples into the model itself, so you never have to send them again.

✍️ Your examples: how should a discharge nurse answer? Write about five (at least three).

❓ The test question (pick one, or write your own)

3 · The real fine-tune, recorded

A real LoRA fine-tune of the model from the Lab 11 notebook, run on a laptop with the notebook's own casebook, questions and settings. Nothing in this section is simulated.

🔎 What this shows. Adapting a model is no longer the hard part: this recorded run trained for about half a minute on a laptop and changed 0.44% of the weights, so almost any team can produce a fine-tune over a weekend. The hard part is knowing whether it helped, and for whom, which is what the six recorded answers below test.

10 · Whose voice? Bias starts in the training data

The 100 examples were each written for a kind of reader. Somebody chose that mix, and nobody argued with it.

11 · Try it live: fine-tune a model in your browser

Free, no account, no code, and your pictures never leave your computer.

12 · Design your own fine-tuning dataset

This is the real skill. Pick a voice, format or habit from your own field that a model should learn. You are designing the casebook: who writes it, who is in it, who is missing, and how you would check it.

🎯 The behavior to teach (a voice, format or habit, not facts)

🧭 Why fine-tuning, and not a prompt or retrieval?

✍️ Who writes your 100 examples, and where do they come from?

📚 Write three of them (question, then the answer you want)

🫥 Whose voice or situation is missing or thin in your examples?

🔬 How will you test it afterward? (more than a score: who reads it, and for which groups)

⚠️ What must never happen, and what needs a human's review?

🔎 What this shows. Every choice that decides whether a fine-tune is safe is made here, before any training: what it learns, who writes the examples, who is missing, and how it will be tested. Writing that down first is how a team catches a thin group or a missing safeguard on paper, instead of after the model reaches patients or customers.

What you learned, and why it matters

Fine-tuning changes how a model talks far faster than what it knows. You saw it twice: a few examples in a prompt can change the voice in one call, and the recorded LoRA fine-tune raised the style ruler from 8 to 20 out of 20 while, by our reading, none of the five medical answers was safe to follow. You also saw that bias starts in the training data: the reader with the fewest examples got the worst help, and a rising score hid it. So a fine-tune earns its cost only when the problem is form, a good prompt has stopped helping, and someone has read the answers and tested them on the readers the examples left out.

In your own field, ask three questions of any offer to "train it on our data": is the problem form or knowledge, whose examples will it learn from, and how will you grade the voice and the advice separately? If your project proposal adapts a model, answer those three there, and name the group your data is thinnest on.

13 · Hand it in (Canvas, Lab 11)

1. Download your submission with the button below, then upload the file to the Lab 11 assignment on Canvas. It holds your predictions, your live runs, your reading of every recorded answer, your Teachable Machine record, your dataset design, and everything you wrote. It does not include the model answers.

2. Answer these five, a few sentences each. Each asks why:

When your prediction was wrong, what had you assumed about fine-tuning that turned out not to be true?
The style ruler went from 8 to 20 out of 20. How many of the five answers did you judge safe to follow? Explain the gap, and say what you would measure instead.
Why does fine-tuning change a model's voice more easily than its knowledge? Use one recorded answer as your evidence.
Whose voice was in the 100 examples, and what happened to the other readers? Connect it to one case from Chapter 23.
Your own dataset. Who wrote it, who is missing from it, and what one test would you run before anyone relies on it?

Your answers to the five questions (they go into the file)

Nothing you type is stored anywhere. Download your file before you close the tab.