Part 2/11:
Fine-tuning is a form of transfer learning, originally developed for image models but now pervasive in NLP. Its primary purpose is to teach a model a new task, not to instill it with new knowledge. Think of transfer learning as tweaking a guitar to improve its performance; it doesn't give you a new instrument. You’re adjusting an existing system rather than replacing its core.
The common misconception is that fine-tuning allows the model to absorb vast amounts of new information—like adding Wikipedia knowledge—so that it can answer queries based on that data. However, this is an oversimplification. Fine-tuning does not teach the model new facts; it teaches the model a pattern or task—like responding to questions in a specific style or format.