Model Customization Decision Framework

Model Customization Decision Framework

Start from a measured failure. Use prompting for instruction, structure, examples, and supplied context. Use retrieval when answers require changing external knowledge with provenance and access control. Consider fine-tuning when a stable task or behavior remains weak after prompt and system design and suitable labeled data exists. Consider distillation when a smaller model must reproduce defined behavior and the teacher, data, license, evaluation, and safety boundaries permit it.

Methods can be combined, but each adds data, evaluation, operations, security, privacy, and license obligations.

Choose the smallest intervention that passes a frozen evaluation set and operating constraints.