Fine-tuning
Continuing to train an existing model on your own examples so it adapts to a task, style or format. Cheaper and faster than training from scratch; still a GPU cost you pay once per run.
What it is
You take a pre-trained base model, feed it a few hundred to a few hundred thousand input/output pairs, and run a handful of training passes (epochs). The weights shift toward your data. The result is a new model you host yourself or upload to a provider that hosts fine-tunes.
Two big families: full fine tuning (update every weight - expensive) and parameter-efficient methods like lora / qlora (update a tiny add-on - cheap). Most people mean the latter.
Why it matters
Fine-tuning is an alternative to a long system prompt + few-shot examples. Done well it lets a smaller, cheaper model match a bigger one on your narrow task - which changes your per-request cost per million tokens for good, not just once.
Done for the wrong reason (adding knowledge - use rag instead) it burns GPU time for little gain.
Cost & infrastructure impact
Cost = GPU-hours for the run + storage of the resulting weights + (if hosted) a per-token serving premium. A lora run on a 7-8B model is often a few dollars to low tens of dollars of GPU time; a full fine tuning of a 70B model is hundreds to thousands.
Model your run with the fine-tuning cost calculator before renting anything.
Related concepts
Use it in Obolith
FAQ
Fine-tuning or RAG?
RAG to give the model facts it should look up. Fine-tuning to change how it behaves - tone, format, following a niche instruction style. They combine well.