Part 6/11:
In reality, fine-tuning only "unfreezes" a small portion of the model, leaving most of it unchanged. It does not upload new facts into the model’s "knowledge base." Instead, it adapts the model to recognize patterns related to a specific task.
This misunderstanding leads people to expect fine-tuning to serve as a knowledge store, but as David points out, it cannot reliably replace dedicated retrieval systems.
The Limitations of Fine-Tuning in Knowledge Representation
Confabulation and hallucination—where models invent facts—are inherent issues.
Models do not possess a true epistemology (a theory of knowledge).
They do not understand why or what they know.
Fine-tuning does not improve a model’s capacity to discern factual correctness.