AI EncyclopediaTechniques & methods
Fine-Tuning
Further training a pre-trained model on a smaller, labelled, task- or domain-specific dataset so it adapts its internal weights to that task - as opposed to prompting, which leaves the model unchanged.
In plain English
If a model is a generalist that has read most of the internet, fine-tuning is like sending it back to school on your specific subject - your support tickets, your legal documents - so it gets noticeably better at that one thing. It genuinely changes the model rather than just phrasing a clever question to it.
Technical explanation
Initial LLM pre-training is self-supervised, over unlabelled text. Fine-tuning is a supervised process: labelled examples specific to a task or domain are used to update the model's weights via further gradient-based training, aligning outputs with the target dataset. This is distinct from prompt engineering, which leaves model weights untouched and instead optimises the input given to the model at inference time.
Why it matters
Fine-tuning is how a general-purpose foundation model becomes a specialised product - a customer-support model that knows a company's own tone and policies, or a coding model tuned on a specific codebase. It is also comparatively costly, which is why most teams test prompting and RAG first.
Real-world example
OpenAI's InstructGPT, the direct predecessor to ChatGPT's approach, combined supervised fine-tuning on human-written demonstrations with a further RLHF stage - fine-tuning and RLHF used together rather than as alternatives.
Common misunderstanding
That fine-tuning is the first thing to reach for when a model 'doesn't know enough' about your data. For most knowledge gaps, RAG is cheaper, faster to update and easier to audit; fine-tuning is better suited to changing a model's style, format or task behaviour than to adding facts.
Something wrong here?
Every entry is hand-researched and hand-written by Nakoda. If a fact is stale, a source has changed or a definition needs sharpening, tell us and we will check it.