← TKF talk · Slide 3 of 18

What is fine-tuning

What is fine-tuning
← PreviousNext →

Fine-tuning is the first thing everyone reaches for. The pitch sounds magical: take a general AI model, feed it a folder of your documents, and watch it become an expert on your domain. The model "learns" your stuff. That's the promise. The reality is more interesting.

✓ WHAT IT ACTUALLY CHANGES - Style and format The model picks up your tone, your vocabulary, your sentence structure, your formatting habits. Show it a thousand policy memos, it starts writing memos that sound like yours.

✕ WHAT IT DOESN'T CHANGE - Facts and recall It doesn't reliably memorize the contents. Ask it about last quarter's budget memo, it might invent a number. The model learned the shape of your writing, not the data in it.

🔧 I didn't know this when I started. I built a whole pipeline to automate fine-tuning the dental policies. Six runs. Three model architectures. The next six slides are what each run actually produced. »»»