Fine-tuning: adapting an AI model to a specific job
Fine-tuning means continuing to train an already trained model on your own examples so it picks up a specific style, format or skill. It is a developer tool: if you simply use an assistant like ChatGPT or Claude, you will almost never need it.
Data as of October 22, 2026
Key figures
| Level | Intermediate |
|---|
The plain-English definition
A large language model first learns from enormous amounts of text. That stage is called pre-training, and it leaves the model knowing a bit about everything but nothing specific about your business. Fine-tuning takes that model and keeps training it on a much smaller dataset built for one job: pairs of “request → correct answer”. The model’s parameters are nudged slightly, and you end up with a new, specialized version.
Hugging Face’s documentation sums it up well: fine-tuning is identical to pre-training except you don’t start from random weights, and it needs far less compute, data and time. The approach took hold in natural language processing with work such as ULMFiT, published in 2018 by Jeremy Howard and Sebastian Ruder.
A concrete example
An insurance company receives customer emails all day and wants to route them automatically: claims, cancellations, address changes, complaints. With a prompt alone, the model sometimes hesitates between two categories or invents a label of its own. After fine-tuning on emails its staff have already sorted, it knows the in-house categories and always returns the same output format, ready for the ticketing system to read.
Why it matters
Fine-tuning makes behavior more predictable and lets you use shorter prompts, which means fewer tokens sent on every call. It also lets you specialize a smaller, cheaper model on a narrow task. Techniques like LoRA train only a small fraction of the weights, which makes the process feasible on modest hardware.
The landscape is shifting: as of October 22, 2026, OpenAI’s documentation states that it is winding down its fine-tuning platform, which no longer accepts new users. The practice remains widespread with open-weight models, which you can browse on our open-source models page.
Common myths and mistakes
- “It’s the best way to add my documents.” Rarely. For knowledge that changes, RAG is easier to maintain: you update the documents, not the model.
- “My assistant fine-tunes itself when it remembers me.” No. An assistant’s memory adds context; it does not change the model.
- “More examples always help.” In practice, the quality and consistency of examples matter more than their number.
- “The tuned model stays just as versatile.” Not always. Heavy specialization can hurt performance on other tasks, so evaluate before and after.
Related terms: pre-training, LoRA, RAG, distillation and open-weight models.
Related terms
- Pre-training · The first training phase of a language model: it learns to predict the next piece of text across a huge corpus, with no specific task.
- LoRA (Low-Rank Adaptation) · A fine-tuning method that freezes the model and trains only small added matrices, making it cheap to adapt a large language model.
- RAG (retrieval-augmented generation) · A technique that first retrieves relevant passages from your documents, then hands them to the model so its answer is based on them.
- Knowledge distillation · A technique that trains a small student model to imitate the outputs of a large teacher model, so it becomes lighter and cheaper to run.
- Open-weights model · An AI model whose trained parameters (the weights) can be downloaded, so you can run it on your own hardware.
- Model parameters · The internal numbers a model tunes during training; how many it has indicates its size, not directly how good it is.
Try it
Full glossary
Frequently asked questions
Does fine-tuning teach a model new facts?
A little, but that is not its strength. It mainly changes behavior: tone, format, labels, the way answers are written. For information that changes often, such as a product catalog or internal docs, RAG is usually the better fit because it supplies the documents at question time.
Can I fine-tune ChatGPT or Claude from the app?
No. Consumer apps never modify the model: memory and custom instructions only add context to your chats. Fine-tuning happens through an API or developer tooling, very often on open-weight models.
Does OpenAI still offer fine-tuning?
As of October 22, 2026, OpenAI's documentation says it is winding down its fine-tuning platform: it is closed to new users, and eligible existing customers can create training jobs until January 6, 2027. Check the official page before building a project on it.
How much data do I need?
Far less than training a model from scratch, but the examples must be clean and consistent. Contradictory or sloppy examples can make results worse rather than better.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.