A support team pastes the same twelve-paragraph instruction block into a chatbot before every single request: “Answer in our brand voice, never promise refunds, always format replies like this.” It mostly works, until it doesn’t, and the prompt keeps growing. That is usually the moment someone asks whether the model could simply be taught these habits once. The answer is AI fine-tuning, and it is both more useful and more limited than the hype suggests.
What Fine-Tuning Means
A large language model starts life as a general-purpose system trained on a vast amount of text. If you want to know how that first stage works, our explainer on how large language models work covers it. Fine-tuning is a second, much smaller training round. You take the finished model and continue training it on your own examples, so its internal weights shift toward the behavior those examples demonstrate.
OpenAI’s model optimization guide describes the idea as training a base model on inputs and outputs you supply so it handles a task more consistently. The same guide is now labeled a legacy page and describes OpenAI’s own fine-tuning platform as being wound down for new users, a reminder that vendor offerings in this area change quickly. The technique itself is not tied to one company, though, and open-weight models can be tuned on your own hardware or rented GPUs.
Fine-Tuning vs Prompting vs RAG
These three approaches solve different problems, and mixing them up is the most common mistake.
Prompting changes what you tell the model on each request. It is instant and free to iterate on, and OpenAI’s guide says plainly that the prompt engineering process “may be all you need” for many use cases. If you have not yet worked through prompt engineering tips, start there.
Retrieval-augmented generation changes what facts the model can see. It fetches relevant documents at request time, which makes it the better choice for knowledge that changes often, such as product catalogs or policies. Our guide to retrieval-augmented generation explains how that pipeline fits together.
Fine-tuning changes how the model behaves. It suits things like a consistent output format, a particular tone, or a narrow classification task where you have many correct examples. It is a poor way to teach a model new facts, since those facts are baked into the weights and go stale the moment they change.
Common Types of Fine-Tuning
Supervised fine-tuning is the standard form: you provide example prompts paired with the ideal responses. According to OpenAI’s guide it works well for classification, specific output formats, and correcting instruction-following failures. Preference-based methods such as direct preference optimization instead show the model a better and a worse answer to the same prompt, which helps with tone, style, and what to emphasize in summaries.
Full fine-tuning updates every parameter in the model, which is expensive for large models. Parameter-efficient methods avoid that. The best known is LoRA, introduced in the paper LoRA: Low-Rank Adaptation of Large Language Models. It freezes the original weights and trains small added matrices instead, and its authors reported cutting trainable parameters by roughly 10,000 times compared with fully tuning GPT-3 175B. That is why hobbyists can now customize models on a single consumer GPU.
What a Fine-Tuning Project Looks Like
The workflow is less glamorous than the concept. You collect examples, clean them, format them (OpenAI’s guide mentions JSONL files), run a training job, and then evaluate the result against held-out test cases. Most of the effort goes into the first and last steps. The guide stresses that test data should be representative of real-world inputs, because a tuned model that looks great on tidy samples can still fail on messy production traffic.
Quality beats quantity here. A few hundred carefully reviewed examples that show exactly the behavior you want will usually serve better than thousands of inconsistent ones, because the model learns the inconsistency too.
When Fine-Tuning Is Worth It
It tends to pay off in a few situations. Your prompt has grown so long that it is costing you in latency and token fees. You need a smaller, cheaper model to match a bigger one on a single narrow task, which connects to the trade-offs in our piece on small language models. Or you need very consistent formatting that prompting cannot reliably deliver.
It tends not to pay off when your data changes weekly, when you have only a handful of examples, or when you have not yet tried a well-written prompt. Fine-tuning also carries a maintenance cost: when the base model is replaced or retired, the tuned version may need to be rebuilt.
FAQs
Does fine-tuning teach a model new facts?
Not reliably. It is better at changing style, format, and task behavior than at adding knowledge. For facts that change or need to be traceable to a source, retrieval is the more dependable option.
How much data do I need to fine-tune a model?
It depends on the task and the method. OpenAI’s guide gives no fixed minimum, and practitioners generally start with a small, high-quality set, evaluate, and add examples only where the model still fails.
Is fine-tuning the same as training a model from scratch?
No. Training from scratch builds a model from random weights and costs enormous amounts of compute. Fine-tuning starts from an already-trained model and adjusts it with a comparatively small dataset.
Can I fine-tune a model on my own computer?
For small and mid-sized open-weight models, yes, particularly with parameter-efficient methods like LoRA. Very large models still need rented cloud GPUs.
Should I try prompting before fine-tuning?
Yes. Prompting is faster to test and costs nothing to change, and many tasks never need anything more. Fine-tune only after a good prompt and, where relevant, retrieval have fallen short.
Discussion about this post