The Hidden Obstacles Of Domain Adaptation In Llms

Customizing an AI model involves balancing new expertise against capabilities it already has. In a June 2024 overview of domain adaptation, Arcee AI describes how further training on company data can make a model forget some of its earlier skills, a problem it calls catastrophic forgetting. It also says customer data is often noisy, repetitive and templated, which makes it hard for the model to generalize to requests that do not look like the training examples.

Supplying relevant documents with each question avoids changing the model's weights, but introduces another problem: finding the right passages. Arcee describes the resulting work of parsing documents, dividing them into searchable sections and adapting the separate search model that picks them. The choice therefore shifts where customization happens, from training the answering model to building the information pipeline around it. Both are Arcee's account of difficulties it has seen, not measurements comparing the two approaches.