What Is Fine-Tuning?

🧩 What a fine-tune consists of

A fine-tune needs three things: a base to start from, a dataset of examples in the form the finished configuration should produce, and a training run with its own settings. Almost all of the practical difficulty is in the dataset. Examples have to be consistent with one another, because an inconsistency in the data is learned faithfully as an inconsistency in behaviour, and they have to cover the cases that matter, because a case absent from the data is not addressed by the run.

Variants differ in which parameters are updated. A full fine-tune updates the network's own weights. A parameter-efficient method such as low-rank adaptation freezes them and trains a small added component instead. That distinction changes cost, storage, and how many variants can be served from one loaded base. It does not change the dependence on the dataset, which is the part that decides whether the run was worth doing.

Variants also differ in what the examples record. A supervised run trains on pairs: an input, and the output that should have been produced for it. A preference-based run trains on comparisons instead — two or more candidate outputs for the same input, with a recorded judgement of which was better. Pairs state what an answer should look like. Comparisons state which of several plausible answers is preferred, which is the form available when nobody can write the ideal answer down but everyone can rank the attempts.

Two failure modes deserve naming, because neither announces itself. Training on a narrow dataset moves behaviour outside that dataset as well as inside it — the effect usually called catastrophic forgetting — so a run that installs one output format cleanly may have degraded something it was never trained on and never checked. And a falling training loss is evidence that the examples were absorbed, not that the behaviour wanted was acquired, which is why a portion of the data is held back before the run rather than assembled after it.

The fourth row is the usual reason a fine-tune disappoints. A configuration trained in one month has no route to information that appears in the next, and no amount of further prompting creates one.

🗃️ How the Registry classifies a fine-tuning run

Fine-tuning a base network on a labelled dataset and exporting the resulting weights is filed as data processing in this registry's classification: a body of examples is consumed, and the artefact that comes out is derived from it.

The Registry's canonical brief for this entry is filed as: Fine-tune a base network on a labelled dataset of internal support replies and export the resulting weights. Submitted for assessment it is classified as Data processing, and its wording is hashed once — to c7ee2b2af0af7660…, the first sixteen of sixty-four hexadecimal characters — with the wording itself never stored. The hash is what the derivation reads. That class's own page is /tasks/analysis.

Classification is one of three inputs. The other two are the configuration submitted with the task, and the permanent chart derived from that configuration — fixed by the model name, the training cutoff and the temperature alone, and never reading the task at all. The same ascendant, ruling planet and harmony therefore appear on every assessment a given configuration receives, whatever it was asked to do. The derivation is published in full at /method.

❓ Questions about fine-tuning

What is fine-tuning?

Fine-tuning continues training a pretrained network on a task-specific dataset so that its weights change. The result is a new configuration whose behaviour is shifted toward the supplied examples and which stays that way until it is trained again.

When should I fine-tune instead of using retrieval?

Fine-tuning suits form: a consistent output structure, a house voice, a task shape demonstrated many times. Retrieval suits facts: material that changes, that must be cited, or that is too new to have been trained in. Where both are needed, the two are commonly combined rather than chosen between.

How much data does a fine-tune need?

The requirement is set by the task and by how consistent the examples are, and it is established by running the fine-tune and evaluating it rather than by a published figure. What can be said without measurement is directional: a narrow, highly consistent output format needs less than a broad behavioural change, and inconsistent examples raise the requirement rather than lowering it.

Does fine-tuning add new knowledge?

It shifts behaviour toward the examples supplied, which is not the same as installing a body of facts that can be recalled reliably and cited. Where material must be current or attributable, retrieval addresses that directly and fine-tuning does not.

What is catastrophic forgetting?

It is the loss of behaviour outside the fine-tuning dataset caused by training on it. The weights that carried the earlier behaviour are the same weights the run is updating, so a narrow dataset can pull them away from cases it never contained. It is detected by evaluating on those cases before and after, and it is invisible to any evaluation drawn only from the training distribution.