Training
Fine-tuning
Also known as: model adaptation
Further training of an already-trained model on a smaller, specific dataset so it specialises in a particular task, style or domain.
Fine-tuning is taking an AI that already knows a lot and giving it extra, focused training so it becomes really good at one particular thing, like your company's tone of voice, or a specialised medical or legal task. Instead of building a model from scratch (hugely expensive), you start from a capable one and adjust it with a smaller set of examples.
Fine-tuning continues the training of a pretrained model on a smaller, curated dataset, nudging its internal settings toward a specific task, domain or style. The base language model supplies broad competence; fine-tuning specialises it.
When to use it (and when not)
Fine-tuning is worth it when you need a consistent style, a narrow task done reliably, or knowledge of a specialised domain. For simply supplying facts or documents, prompting or retrieval is usually cheaper and easier to update. A good rule: use prompting/retrieval for knowledge, fine-tuning for behaviour and form.
Instruction tuning
A special, important case is instruction tuning: fine-tuning a raw model on examples of instructions and good responses so it becomes helpful and follows requests, the step that turns a raw text-predictor into a usable assistant.
Fine-tuning updates the weights of a pretrained network on a task-specific distribution, trading generality for specialisation. Full fine-tuning updates all parameters and is costly to train and serve; parameter-efficient methods (PEFT), notably LoRA and its quantised variant QLoRA, freeze the base weights and train small low-rank adapters, cutting compute and storage by orders of magnitude while approaching full fine-tuning quality (Hu et al. 2021).
In the LLM pipeline, alignment is itself fine-tuning: supervised fine-tuning (SFT) on demonstrations followed by preference optimisation, RLHF or its direct-optimisation successors such as DPO, is what produced instruction-following assistants (Ouyang et al. 2022). Practical risks include catastrophic forgetting of general capability, overfitting to small datasets, and data-quality sensitivity. The standard decision heuristic contrasts fine-tuning with in-context alternatives: retrieval-augmented generation for updatable knowledge versus fine-tuning for durable behavioural change, with the two often combined.