What is Adapter Layers?

Adapters address the cost of fine-tuning. Rather than updating every parameter in a large model, small layers are inserted at intervals and only those are trained, which reduces compute requirements by orders of magnitude and allows several specializations to share one base model. The approach is widely used where an organization wants domain-specific behaviour without the expense of a full training run. It also raises a governance question that is often overlooked: an adapter trained on proprietary material encodes that material, so where the adapter is stored and who can access it becomes an intellectual property question rather than a technical one. Centralpoint's position is to keep organizational knowledge in prompts and skills held in the organization's own environment — readable, versioned and portable across models — rather than committed to weights that are difficult to inspect and tied to a specific base.