Fine-Tuning
Continuing training of a pretrained model on a smaller, task-specific dataset to specialize it.
Definition
Fine-tuning takes a pretrained model and continues training it on a focused dataset so it performs better on a specific task or domain. It is a form of transfer learning that adjusts existing weights rather than learning from scratch.
An alternative to fine-tuning is in-context learning, where examples are placed in the prompt and no weights change at all. The two approaches trade off: fine-tuning bakes in behavior durably, while prompting is flexible and immediate but consumes context on every request.
Parameter-efficient methods such as low-rank adaptation have changed the economics of specialization: instead of copying and retraining a whole model, one trains a small set of added weights, so many task-specific variants can share a single base. This makes it practical to maintain specialized behaviors without the storage and compute of full copies, and it reduces the risk of degrading the base model's general ability.
Techniques
- Full fine-tuning: update all parameters (most costly).
- Parameter-efficient methods (LoRA, adapters): update a small added set.
- Instruction tuning: teach a model to follow task descriptions.
Why it matters
Fine-tuning delivers specialized performance without the cost of pretraining, and parameter-efficient variants make it feasible on modest hardware. Its risk is catastrophic forgetting, where the model loses general ability while gaining narrow skill.
Fusion connection
A general surrogate can be fine-tuned on a specific Hyperion operating regime to sharpen accuracy where the design study is concentrated.