Parameter-efficient fine-tuning (PEFT) adapts a model by training a small number of additional or decomposed parameters while keeping most base weights frozen. It can reduce training cost and simplify version management compared with full fine-tuning. section 9.3.3
PEFT is one of several model-customization approaches:
- LoRA and adapter modules are common PEFT techniques that add or decompose trainable parameters. section 9.3.3
- Prompt learning trains continuous prompt, prefix, or soft-prompt parameters while keeping the main model frozen. section 9.3.3
- Alignment optimization adjusts behavior using preference or policy signals, rather than defining customization by which parameters are trained. section 9.3.3
Fine-tuning for an enterprise use case can draw on domain data such as support tickets or internal task demonstrations. It also requires dataset governance, training infrastructure, validation, and rollback plans. section 9.3.3







