What does Fine Tuning mean?

Fine tuning means continuing to train an already pre-trained model on your own, considerably smaller dataset. It changes the model’s weights, the numerical values that determine how strongly an input influences the output. Fine tuning differs from prompting in that it alters the model itself permanently, not just its input.

A model with eight billion parameters holds just as many such weights. Full fine tuning adjusts all of them, which requires computing power and memory on the scale of the original training run. More common are methods such as LoRA: they leave the base model untouched and train only small additional matrices, cutting the resource requirement to a fraction. Training runs on pairs of input and desired output, usually a few hundred to a few thousand examples.

Fine tuning pays off wherever a model needs to produce a fixed format, a particular tone, or a set classification, and examples for it already exist. Typical cases are writing in a defined house style, recognising domain-specific terms, and sorting items into a fixed set of categories. For current factual knowledge, it is the wrong tool.

Its advantage over lengthy instructions in the prompt lies in input length. Format and wording are baked into the model after training and no longer need to be sent with every call. A smaller, fine-tuned model can thereby often match the quality of a much larger one.

Fine tuning does not add new domain knowledge to the model; that is what runtime access to verified sources is for. Every adapted model version needs its own identifier, so it can later be traced which version produced a given output.