Fine-tuning in AI means continuing the training of an already pretrained model on additional, task-specific examples. The result is a new version of the model that handles a particular task, vocabulary or style better than the base model. The word is also used in physics and theology, which this entry does not cover.
How does fine-tuning work?
IBM describes fine-tuning as adapting a pretrained model with a smaller, task-specific dataset and as a subset of transfer learning, which reuses existing model knowledge instead of training from scratch. Common approaches are:
- Full fine-tuning: updates all parameters. It is effective but needs a lot of computing power.
- Parameter-efficient fine-tuning (PEFT), for example LoRA: changes only a small part of the parameters, which reduces the effort.
- Instruction tuning: teaches a model to follow instructions using pairs of prompts and desired responses.
- RLHF: human feedback steers qualities such as helpfulness or tone.
IBM also lists risks. A model can forget earlier knowledge (catastrophic forgetting) and can overfit on small datasets.
Fine-tuning, prompting and RAG
Prompting gives the model instructions and context with every request. Retrieval-augmented generation (RAG) fetches matching documents at the time of the request. Fine-tuning changes the model itself. Databricks and Progress both name data that changes often and the need for source citations as reasons to prefer RAG. Fine-tuning fits when a consistent style or a specialized domain is required. The two can be combined. According to OpenAI’s documentation (as of October 2026), OpenAI is winding down its fine-tuning platform and new users cannot access it.
What does it mean for SEO and AI visibility?
A website cannot be added to someone else’s model through fine-tuning. Providers or companies with their own data run it. For websites, the routes into AI answers matter more, namely training data and retrieval at the time of the request (assessment).
Example
A support team fine-tunes a model on several hundred past replies so it answers in the house tone and a fixed ticket format. Prices and delivery times would go stale inside the model, so the system passes them in per request via RAG.
FAQ
What does fine-tuning mean? Training an existing model further on specific examples so that it fits one task or style.
Is fine-tuning the same as training? It is a second training phase on an already trained model, usually with much less data than the original training.
Sources
- What is fine-tuning? - IBM Thinkibm.com
- What is fine-tuning? - Databricksdatabricks.com
- RAG vs fine-tuning - Progress (secondary source)progress.com
- Model optimization (fine-tuning) - OpenAI Docsdevelopers.openai.com
Last updated: