Fine-tuning
Specializing an AI model for a specific task or style.
Teaching an old model new tricks without retraining from scratch.
The idea
A model is pre-trained on billions of words. Fine-tuning means showing it hundreds or thousands of your own examples (your style, your data, your format) and letting it adjust its dials slightly to match.
It's cheaper than training from scratch but more intensive than a one-off prompt.
When to fine-tune
Fine-tune when your task is highly specialized, like legal language, your brand's tone, or a niche domain. Fine-tune when you will run the model thousands of times, so the per-call savings add up. And fine-tune when you want consistency across many generations.
Start with prompting and RAG first. Fine-tuning is premature if a careful prompt or two examples get you there.
Pick what you are building
Reach for the cheapest that works
- prompt itcheapest
- add retrievalmore effort
- fine-tunemost upfront cost
What fits this task
You want your support bot to sound warm and on-brand.
prompt itA few lines of instruction and one example reply set the tone. Nothing about the model needs to change, so start here.
Cheapest and fastest to try. You change your words and leave the model alone.
Start at the top of the ladder and stop as soon as something works. A good prompt is free to try, retrieval hands the model facts it never saw, and fine-tuning is the far end. It pays off once the task is narrow and you run it thousands of times, so treat it as a scaling play.
The trade-off
Fine-tuning costs money upfront and takes time. But a fine-tuned model becomes cheaper to run than a flagship for high volume. It pays off once you have real volume, so treat it as a scaling play.