What is fine-tuning?
Continuing a model's training with new examples so its behavior shifts. Small word, big fork in the road: adapt the model, or give it something to read.
Fine-tuning takes a trained model and continues its training on new examples, nudging the weights until the model's behavior matches them. It changes how a model writes, formats and responds; it does not make the model look things up. For knowledge, a document library with retrieval stays better, because it is current, checkable and can be swapped out at any time. Fine-tuning earns its place where style or format needs to be consistent.
More training, on purpose.
A model's weights are the residue of its training data, and fine-tuning resumes that process: the model sees new examples, and the same mechanism that produced the weights in the first place adjusts them slightly to fit. The result is an open-weight model that behaves differently, not one that knows it should behave differently. The distinction matters because behavior learned this way is statistical and invisible; there is no file to inspect, only tendencies, which is why tuning works best for things you can demonstrate in examples rather than state in a paragraph.
Knowledge is the wrong tool here.
The most common misuse is tuning a model on your own documents, hoping it will absorb them. What actually happens is lossy compression: facts sink into the weights without their sources, immune to update and impossible to cite. That is the reverse of what research work needs. Retrieval-augmented generation keeps the knowledge outside the model, in documents that stay current, and the answer can point at the passage it came from. The division of labor is simple: behavior goes in the weights, knowledge goes in the library.
Why it became practical.
Full fine-tuning rewrites every parameter and wanted hardware few people had. Methods such as LoRA train a small adapter alongside the frozen weights instead, which shrinks the compute by orders of magnitude and makes adapting a 7B model a weekend project on consumer hardware. Even so, the honest baseline for local research tools is that they need no tuning at all: a general model that reads carefully, pointed at a good library, already does the work most people imagine they would need a custom model for. And when the goal is not a custom model but a smaller one that behaves like a bigger one, that job belongs to distillation.