Fine-tuning

Fine-tuning is the operation of specializing an already pre-trained model, such as a foundation model, by having it undergo additional training on a set of targeted data, specific to a particular task or domain. Rather than training a model from scratch, which would require colossal resources, one capitalizes on the general knowledge already acquired and directs it toward a precise use, for example the analysis of legal contracts or medical diagnosis. This technique has democratized the use of artificial intelligence, by making it accessible to many organizations to create models adapted to their needs at a reasonable cost. It carries, however, risks of its own that directly concern risk management and insurance. Fine-tuning can reintroduce or amplify biases present in the adaptation data, unintentionally degrade the safety guardrails carefully put in place by the base model's designer, or expose confidential data used for training, which a poorly protected model can subsequently regurgitate. Fine-tuning thus shifts part of the responsibility toward the organization that adapts the model, blurring the apportionment of responsibilities between the designer of the base and the specializing user.

A company fine-tunes a foundation model on its own files to automate claims handling. A poorly controlled fine-tuning can lead the model to regurgitate, in its answers, confidential data present in the training files.

fine-tuning, réglage fin, ajustement fin, affinage