K-Culture GlossaryㅍTechnical words in the news
Fine-tuning
Giving a already-built model extra training for a specific job — more like altering a ready-made suit than tailoring a new one.
In plain words
Fine-tuning means taking a model that's already been built and giving it extra training for a specific purpose. If building a new model from scratch is like having a suit custom-tailored, fine-tuning is like altering a ready-made suit — you buy a good jacket (a general-purpose model) and just shorten the sleeves to fit your body (your company's work).
The method is feeding the model prepared example data. Fine-tune it on thousands of medical consultation records, and you get a model steeped in medical language and knowledge. Fine-tune it on your own company's customer service logs, and you get a model that answers the way your company does. It's far cheaper than training a whole new model from the ground up.
In articles, this comes up in two contexts: companies announcing they've "built their own fine-tuned model" as an adoption case study, and news of derivative models fine-tuned from open-source models (like "a Llama-based Korean-specialized model") being released. It's worth remembering the difference from RAG, where changing the source data changes the answers — fine-tuning instead makes the knowledge "second nature" to the model itself.
How it shows up in the news
"By fine-tuning an open-source model on our own data, we cut costs to one-tenth." — a classic adoption story about choosing an altered ready-made suit over a big tech API.
Fine-tuning vs. Retrieval-Augmented Generation: what's the difference
Does it retrain the model's brain?, Does it hand over reference material before answering?
People use the phrase "we trained the AI on our company data" for both, but the methods are completely different. One retrains the model itself to change its underlying behavior, while the other leaves the model untouched and simply fetches relevant documents to hand over at query time. When a company says "let's build an internal chatbot," what they usually need is the latter.
| Aspect | Fine-tuning | RAG |
|---|---|---|
| What changes | Retrains the model itself | Model stays the same. What changes is the material handed to it |
| What it's good at | Internalizing tone, format, and a feel for a specific domain | Answering accurately based on up-to-date facts and internal documents |
| When the data changes | Requires retraining. Costs time and money | Just swap the documents and it's reflected immediately |
| Can it cite sources? | Cannot point to where it learned something | Can show the retrieved documents as evidence |
| Analogy | Training an employee | Placing reference materials on the employee's desk |
Rule of thumbIf the issue is tone and format, it's fine-tuning; if the issue is what the model knows, it's Retrieval-Augmented Generation.
Fine-tuning vs LoRA: what's the difference
Retrains the entire model, Trains only a small added piece
LoRA is itself a type of fine-tuning, so asking 'which one should I use' is a bit of a false choice. Fine-tuning is the broad category, and LoRA is a cheap way of doing it. When articles today say a model was 'fine-tuned,' it usually actually means LoRA.
| Aspect | Fine-tuning | LoRA (Low-Rank Adaptation) |
|---|---|---|
| What gets trained | All the numbers inside the model | Only a small added piece |
| Cost | Requires enterprise-grade hardware | Has come down to the point where a personal PC can handle it |
| Output | An entirely new full model is produced | The original stays as is; a single small file is produced |
| Swapping | The whole model has to be replaced | Only the piece is swapped in and out; multiple can be used interchangeably |
| Analogy | Making a new set of clothes | Adding a patch to existing clothes |
Rule of thumbFine-tuning is the goal, and LoRA is a cheap way to achieve that goal — it's not a matter of choosing one or the other.
See also
Stories using this term
No story has used this term yet. New ones attach here automatically.