Fine-Tuning

Level 1

Teaching an AI that already knows a lot to be extra good at one special thing — like a dog that learns a new trick.

Level 2

Taking an AI that's already been trained on general stuff and giving it extra training on specific topics so it becomes an expert.

Level 3

The process of further training a pre-trained model on a smaller, task-specific dataset to adapt it for particular use cases.

Level 4

Continued training of a pre-trained foundation model on domain-specific or task-specific data, often with modified learning rates, to specialize its capabilities while preserving general knowledge.

Level 5

Parameter adaptation of a pre-trained model via gradient updates on a downstream objective — encompassing full fine-tuning, LoRA, prefix tuning, and RLHF, each navigating the plasticity-stability tradeoff differently.

Sources

• NIST Trustworthy and Responsible AI Resource Center: Glossary — National Institute of Standards and Technology. Accessed 2026-07-20.