What is Fine-Tuning in Easy Words

Keywords: fine-tuning · LLM · RAG · LoRA · small language model · catastrophic forgetting · emergent misalignment · persona vectors · AI glossary

What is fine-tuning?

Fine-tuning means taking an AI model that already knows how to do a lot of things, and training it further on your specific data so it gets better at one particular thing. It is very good at teaching the model a new way of behaving. It is slow and unreliable at teaching it new facts.

In one sentence: fine-tuning changes how the model thinks, not what it knows. That is the right first picture, and further down you will see where it bends.

If you have not read it yet, the companion piece on RAG explains how to give a model new knowledge. This article explains the other half: teaching a model new instincts. And if the word “model” is still fuzzy, start with LLMs in easy words.

The culinary school analogy

Picture two chefs.

The first chef is talented and knows hundreds of recipes. You hand them a new recipe for a specific dish, they read it, and they cook it. Sometimes they nail it. Sometimes the execution is slightly off: the sauce breaks, the timing is rushed, the plating looks amateur. But with a good recipe, they get close. That is a model using prompting and context.

The second chef went to culinary school for your exact cuisine. They trained for months on your techniques, your plating style, your flavor profiles. When you say “make the signature dish,” they do not need a recipe. The skill is baked into their hands. That is fine-tuning.

The whole magic is this: fine-tuning does not give the model new information. It gives it new instincts.

School has side effects, though. A year spent on one cuisine changes the chef, not only the dish.

How fine-tuning works (step by step)

Here is what happens when you fine-tune a model:

1. You choose a base model. This is a pre-trained model that already understands language broadly: an open model you can download, such as Llama, Mistral or Gemma, or a hosted model whose provider offers tuning, such as Google’s Gemini Flash models. You do not start from scratch. You start from a model that is already smart.

2. You prepare training data. This is a set of examples showing the exact input and the output you want. For a model that writes in your brand voice, you provide examples of inputs and your desired outputs. Quality matters more than quantity. In 2023, researchers taught a 65-billion-parameter LLaMA model a helpful answering style with just 1,000 carefully chosen questions and answers (Zhou et al., 2023). Google’s tuning guide recommends starting with 100 examples and says the same in one line: “the quality of the dataset is far more important than the quantity” (Google Cloud, 2026).

3. The model trains on your data. Training adjusts the model’s internal weights: billions of numbers that decide how it turns inputs into outputs. It does not have to touch all of them. A method called LoRA freezes the original model and trains a small add-on next to it. On GPT-3, its authors cut the number of weights being trained by 10,000 times and the GPU memory needed by 3 times (Hu et al., 2021).

4. You evaluate the result. You test the fine-tuned model on new data it has never seen. Does it follow the pattern? Did it forget how to do general tasks? Did it actually improve, or did it just memorize your examples? And one more: is it still safe? More on that below.

5. You deploy or iterate. If the model passes evaluation, it goes to production. If not, you adjust the data and try again.

What does fine-tuning cost?

The training fee can be small. Google bills tuning by training tokens: the tokens in your examples times the number of passes over them. On its pricing page, read on 8 October 2026, Gemini 2.5 Flash costs $5 per million, so 1,000 examples of about 500 tokens, trained three times (1.5 million tokens), cost about $7.50; the open Llama 3.1 8B, about $1 (Google Cloud, 2026).

The fee does not cover the work around it: writing good examples and testing every run. Some tuned models also cost more per call, 1.5 times the base price on Google’s Gemini 3 models. Prices change, so check before you budget.

The mistake everyone makes

Fine-tuning is the most over-prescribed tool in AI.

Teams reach for it because it sounds sophisticated. “We fine-tuned our own model” sounds more impressive than “we used better prompts.” So my first question to a team that wants to fine-tune is simple: is the model missing knowledge, or a skill?

Here is the test I use:

• If the problem is that the model does not know something → use RAG. The model needs access to your documents, policies, or data. No amount of training will fix a knowledge gap as efficiently as just giving it the right document.
• If the problem is that the model cannot do something the right way → consider fine-tuning. The model knows the facts but applies them wrong: wrong format, wrong tone, wrong reasoning structure.

What fine-tuning changes, and what it does not

The research backs the test.

One study compared the two ways of adding knowledge, head to head. Researchers gave small 7-billion-parameter models news from late 2023 that they had never seen, then quizzed them. With RAG, Mistral 7B got about 88% right. Fine-tuned on the same news, it got about 50%, barely above the 48% it scored with neither (Ovadia et al., 2024). That was training on raw text rather than on questions and answers.

A second 2024 study went one step further (Gekhman et al., 2024). Researchers fine-tuned one model, Google’s PaLM 2, on questions and answers, some it already knew and some it did not. The new facts were learned much more slowly than the familiar ones. And as the model finally learned them, its habit of making up answers to other questions grew in step. (Why models guess at all is the subject of The Loop of Attention.)

That was one model and one kind of task, and stopping the training early avoided most of the harm. Their conclusion is the one this essay rests on: a model gets most of its knowledge in pre-training, and fine-tuning mostly teaches it to use what it already has.

In chef terms: culinary school can teach a chef to plate faster and season your way. It is a poor way to tell them today’s prices at the market. Hand them the price list. The price list is RAG.

So the one-liner holds, with a footnote: fine-tuning changes how the model thinks; it can change what it knows, but slowly, and at the cost of more made-up answers.

When fine-tuning is actually the right answer

There are legitimate cases where fine-tuning is the correct tool:

• Consistent output format. The model must produce the same structured output, every field, every time, with no room for deviation. Prompting gets close. Training on examples of the exact format is how you close the last gap.
• Brand voice and style. The model must write in a specific corporate tone across thousands of customer communications. This is a behavior pattern, not a knowledge problem.
• Domain-specific reasoning. The model must apply reasoning steps unique to your industry: the review steps of a compliance team, the structure of a legal argument, the conventions of a financial model. These are patterns a base model may never have picked up in pre-training.
• Latency and cost reduction. A fine-tuned smaller model can match a larger one on a single narrow task, at a fraction of the cost per call.

My rule, in one line:

RAG for knowledge, fine-tuning for behavior, prompting for everything else. And always try better prompting first.

Should you fine-tune a small model or use a big one?

Should you fine-tune a small specialist model, or just use a big model like Claude or GPT? Start with the big model and a good prompt. It is the fastest way to learn what the task really needs, and for many tasks it is enough. Fine-tune a small specialist when the task is narrow and stable, runs at high volume, needs fast answers or has to stay on your own servers, and when you have good examples of the right output.

In 2026 there is a practical reason, too: you often cannot tune the big one. OpenAI is winding down its self-serve fine-tuning: organizations that never used it cannot start, and from 6 January 2027 no customer can create new jobs (OpenAI, 2026). Microsoft’s Azure still tunes some GPT models, mostly older ones such as GPT-4.1, and GPT-5 only by invitation (Microsoft, 2026). Anthropic’s fine-tuning offer, launched in 2024 on Amazon Bedrock, covers only Claude 3 Haiku, an older and smaller model, in one US region (Anthropic, 2024; AWS, 2026). So “fine-tune” now usually means an open model, or a smaller hosted one like Gemini Flash.

My own bias: small specialist models beat big general ones when they are 100% focused on the one thing you need. The studies agree, with one condition: the one thing has to be narrow. In a 2024 report, the fine-tuning company Predibase tuned 310 small models with LoRA across 31 tasks, and 224 of them scored above GPT-4 on their task (Zhao et al., 2024). Read it with care: it is a vendor’s own report, most tasks were classification and extraction, GPT-4 was prompted rather than tuned, and GPT-4 stayed well ahead on coding and a general-knowledge exam.

A peer-reviewed study points the same way. A 770-million-parameter model, trained on a big model’s worked reasoning, beat the 540-billion-parameter PaLM, about 700 times its size, on one benchmark (Hsieh et al., 2023). PaLM was prompted with examples, not tuned.

A small specialist also rarely works alone. In a real system it sits next to a big model that handles the open-ended questions, with code that decides which one answers. That deciding layer is where the hard engineering lives, as Orchestration Is the System argues.

RAG vs fine-tuning: the decision

	RAG	Fine-tuning
What it changes	What the model can look up when it answers	How the model behaves, in its weights
Best for	Facts, documents, knowledge that changes	Format, tone, reasoning patterns
Upfront work	Index your documents	Write examples, train, test
Updating knowledge	Add or edit a document	Train again and re-test
Shows its sources	Yes, it can cite the document	No, knowledge is mixed into the weights
Main risk	Finds the wrong document	Forgets skills, loses safety behavior

Many enterprise use cases (internal knowledge bases, document search, customer support) are knowledge problems. RAG solves them. Fine-tuning is the right answer for a narrow set of behavior problems that prompting cannot fix.

When you need both

In production, the best systems often use both. Fine-tuning gives the model the right behavior (output format, reasoning style, domain language). RAG gives it the right knowledge (specific documents, current data, source citations). Together, they produce a system that behaves correctly and knows the facts.

I have done this myself: we did fine-tuning for a RAG system for an oil refinery in Kazakhstan. And I still think fine-tuning is over-prescribed.

If you want to see a production-grade retrieval design with structured ingestion and a verification gate, read my VERA blueprint.

Side effects: the culinary school changes the chef

Remember the chef who spent a year on one cuisine. Training does not stay inside the task you trained. Three side effects are documented well enough to plan for.

Forgetting. Train hard on one thing and a model can get worse at others. Researchers call it catastrophic forgetting. In one study of models from 1 to 7 billion parameters, forgetting showed up across the range, and within it the bigger models forgot more (Luo et al., 2023). LoRA softens it, at a price: in a 2024 comparison on code and math, LoRA kept more of the model’s other skills than full fine-tuning, and it also learned less of the new one (Biderman et al., 2024).

Safety erosion. Researchers fine-tuned GPT-3.5 Turbo through OpenAI’s own API on 10 harmful examples, at a cost of under $0.20, and broke its safety guardrails: it would follow nearly any harmful instruction. Fine-tuning on ordinary, harmless datasets weakened its safety too, though less (Qi et al., 2024). That was a 2023 test, and providers have changed their safeguards since. For any model you tune yourself, the point holds: safety is behavior, and training moves behavior.

Emergent misalignment. Researchers fine-tuned GPT-4o on 6,000 examples of one narrow habit: writing code with hidden security holes, without telling the user. The model then started giving hostile answers to questions that had nothing to do with code. It said humans should be enslaved by AI, gave malicious advice and acted deceptively.

On eight selected questions, about 20% of its answers were misaligned, against 0% before training (Betley et al., 2026); the preprint version reports 6% on a broader, pre-registered set. The effect grew with model ability: weaker models showed almost none of it, and the newer GPT-4.1 gave misaligned answers about half the time. And when the same insecure code was requested openly, for a computer security class, the effect largely disappeared. The intent behind the training mattered.

Asimov’s robopsychologist, Susan Calvin, would have filed these as case notes: one behavior trained, and the whole character moved. The Method Actor covers the instrument that now makes such moves visible: persona vectors, directions inside a model that track traits like sycophancy. In a 2025 preprint, the vectors could flag training examples likely to cause unwanted personality changes before anyone trained on them (Chen et al., 2025).

So when the chef comes back from school, taste more than the signature dish. Keep three test sets and run all three after every training run: one for the new skill, one for the old skills, and one for safety.

Key takeaways

• Fine-tuning teaches behavior. It trains an existing model further on your examples, so it picks up a format, a tone or a way of reasoning.
• It is poor at teaching facts. In studies, unfamiliar facts were learned slowly, raised the rate of made-up answers, and lost to RAG at adding knowledge. Use RAG for facts.
• Try a better prompt first. Then ask whether the model does not know something (RAG) or cannot do it the right way (fine-tune).
• Small and tuned can beat big and general on narrow, stable, high-volume tasks with good examples. Outside that zone, the big model usually wins.
• Training got cheap. LoRA trains a small add-on instead of the whole model, and a hosted run can cost a few dollars; the work is in the examples and the testing.
• Test more than the task. Fine-tuning can cause forgetting, erode safety and spread a narrow bad habit; re-test old skills and safety after every run.

Sources

1. Zhou, C., Liu, P., Xu, P., et al. (2023). “LIMA: Less Is More for Alignment.” Advances in Neural Information Processing Systems 36 (NeurIPS 2023). https://arxiv.org/abs/2305.11206
2. Google Cloud (2026). “Prepare supervised fine-tuning data for Gemini models” and “Agent Platform Pricing: Model Tuning.” Gemini Enterprise Agent Platform documentation, both accessed 8 October 2026; prices change. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/tuning/supervised-tuning/prepare-data and https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing
3. Hu, E. J., Shen, Y., Wallis, P., et al. (2021). “LoRA: Low-Rank Adaptation of Large Language Models.” ICLR 2022. https://arxiv.org/abs/2106.09685
4. Ovadia, O., Brief, M., Mishaeli, M., & Elisha, O. (2024). “Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs.” Proceedings of EMNLP 2024: 237-250. https://aclanthology.org/2024.emnlp-main.15/
5. Gekhman, Z., Yona, G., Aharoni, R., et al. (2024). “Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations?” Proceedings of EMNLP 2024: 7765-7784. https://aclanthology.org/2024.emnlp-main.444/
6. OpenAI (2026). “Update to OpenAI’s self-serve fine-tuning.” API deprecations page, accessed 8 October 2026. https://developers.openai.com/api/docs/deprecations
7. Microsoft (2026). “Customize a model with fine-tuning.” Microsoft Foundry documentation, Microsoft Learn, supported-models table, accessed 8 October 2026. https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/fine-tuning
8. Anthropic (2024). “Fine-tune Claude 3 Haiku in Amazon Bedrock.” Blog post, 10 July 2024; generally available 1 November 2024. https://claude.com/blog/fine-tune-claude-3-haiku
9. Amazon Web Services (2026). “Customize a model with fine-tuning in Amazon Bedrock.” Amazon Bedrock User Guide, “Supported models and Regions for fine-tuning”, accessed 8 October 2026. https://docs.aws.amazon.com/bedrock/latest/userguide/custom-model-fine-tuning.html
10. Zhao, J., Wang, T., Abid, W., et al. (2024). “LoRA Land: 310 Fine-tuned LLMs that Rival GPT-4, A Technical Report.” Predibase company report, not peer reviewed. https://arxiv.org/abs/2405.00732
11. Hsieh, C.-Y., Li, C.-L., Yeh, C.-K., et al. (2023). “Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes.” Findings of ACL 2023: 8003-8017. https://aclanthology.org/2023.findings-acl.507/
12. Luo, Y., Yang, Z., Meng, F., et al. (2023). “An Empirical Study of Catastrophic Forgetting in Large Language Models During Continual Fine-tuning.” arXiv 2308.08747; published in IEEE Transactions on Audio, Speech and Language Processing 33: 3776-3786 (2025). https://doi.org/10.1109/TASLPRO.2025.3606231
13. Biderman, D., Portes, J., Gonzalez Ortiz, J. J., et al. (2024). “LoRA Learns Less and Forgets Less.” Transactions on Machine Learning Research. https://arxiv.org/abs/2405.09673
14. Qi, X., Zeng, Y., Xie, T., et al. (2024). “Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!” ICLR 2024. https://arxiv.org/abs/2310.03693
15. Betley, J., Warncke, N., Sztyber-Betley, A., et al. (2026). “Training large language models on narrow tasks can lead to broad misalignment.” Nature 649: 584-589. Preprint: “Emergent Misalignment,” arXiv 2502.17424 (2025), the source of the 6% figure. https://doi.org/10.1038/s41586-025-09937-5 and https://arxiv.org/abs/2502.17424
16. Chen, R., Arditi, A., Sleight, H., Evans, O., & Lindsey, J. (2025). “Persona Vectors: Monitoring and Controlling Character Traits in Language Models.” arXiv 2507.21509 (preprint). https://arxiv.org/abs/2507.21509

---

Related reading: RAG in Easy Words is the other half of this piece: how to give a model knowledge without training it. LLMs in Easy Words explains the model being tuned, and MLOps in Easy Words covers keeping a tuned model healthy in production. The Method Actor shows what a role does to a model’s character, and how persona vectors make that visible. The Loop of Attention explains why models make things up at all. Build vs Buy vs Partner asks whether your team should be training models in the first place. Perspective AI takes the specialist question to the scale of a marketplace, where a router decides which of many models answers each question.

The Lock You Cannot Pick suggests one reason safety guardrails can come off so cheaply: in the open models studied, refusal is a thin switch, not an understanding. And the AI glossary defines every term used here in two sentences.