Fine-Tuning

The process of further training an existing, typically pre-trained, model on a smaller, more specific dataset to adapt its behavior to a particular task or domain. Fine-tuning creates a State 2 retention relationship between the fine-tuning dataset and the resulting model, distinct from the retrieval-based State 3 relationship a RAG system creates with the documents it looks up.

The operator-side form of tuning, in which an organization trains a base model further on its own examples to produce a customized version of the weights. It can create behavior no prompt or context can, but it is expensive, hard to reverse, and produces a new model artifact the organization must govern, so experts reach for it only after the lighter dials are exhausted.

A subsequent, smaller-scale training process applied to an already pre-trained model, using a narrower dataset to specialize its behavior for a specific use case or to align its behavior with particular guidelines. Two products built on the same pre-trained model can behave very differently after different fine-tuning.

Continuing a model's training on a curated set of domain-specific examples so a particular style, format, or skill is built into the model's weights. It changes the model itself, in contrast to RAG, which changes what the model is shown.

The process of further training an already-trained model on a smaller, task-specific dataset, used to adapt a general foundation model to a specific robot, task, or environment.