Classification Fine-Tuning

This unlocks the ability to make a trained language model answer 'which category?' questions with confidence — spam vs ham, positive vs negative review, which department an email belongs to, whether a transaction looks fraudulent. It is the mechanism behind nearly every 'smart' feature that sorts or flags things. In the path ahead, this is the bridge between a model that only generates text and a model that acts as a decision engine. It also connects directly to your notes on Instruction Tuning (teaching directions vs teaching classification), Catastrophic Forgetting (what you risk when you tune), and LoRA (the efficient way to do it).

A language model, before any fine-tuning, is a machine that predicts the next word. Given 'The movie was', it might continue 'great', 'boring', or 'unexpected'. That is generative — it produces an open-ended continuation. Classification is different: you have a fixed set of possible answers (labels), and you want the model to pick the correct one. For example, given a review, the fixed answers might be 'positive' and 'negative'. You don't want it to generate a sentence; you want it to output 'positive' if the review is good, 'negative' if bad. Fine-tuning for classification is the process of taking that general word-predictor and turning it into a specialised label-picker. You do this by taking the model, attaching a small new part on top (a classification head), and then showing it many examples of inputs together with their correct labels. Over many examples, the model adjusts its internal numbers so that the head at the top produces the right label for each input. The key idea is that the base model already understands language extremely well — it knows that 'amazing' is a positive word, 'terrible' is negative. Fine-tuning just teaches it to use that understanding in a specific way: to compress the whole input into a single choice. The input might be a whole paragraph; the model sums up its understanding into a final vector at the last token, and the head reads that to decide the label.

Think of a seasoned librarian who is brilliant at understanding text but has never been trained to sort books into categories. She can read a synopsis, follow the plot, know the characters. But if you ask her to put each book into a box labelled 'fiction' or 'non-fiction', she has no idea what system to use. Now, you ask her to read 10,000 books along with their correct categories. After a while, she learns what features in a book's synopsis tend to mean it's fiction — invented names, a quest, dialogue — and what features tend to mean non-fiction — a thesis, references to studies. She doesn't lose her reading ability; she simply adds a new skill of labelling. The librarian is the base language model. The new skill of labelling is the classification head. The 10,000 labelled examples are the fine-tuning dataset. After training, you can hand her a book she has never seen, and she will file it correctly with high confidence. Now, where does the analogy break? A real librarian could reason about the label system and transfer it to novel situations immediately; a fine-tuned model can only classify in ways similar to what it saw. Also, a librarian does not change her brain's wiring when she learns to classify — but a model does, by adjusting millions of numbers (weights). And crucially, the librarian reads the whole book; a language model, in practice, only sees a limited window of context (for example, 512 or 1024 tokens). The model does not 'read' the entire input like a human; it mathematically processes the token sequence to produce a summary vector at the end.

Classification fine-tuning is the process of taking a pretrained language model, attaching a small classification head on top of the final token's representation, and training the whole thing on labelled examples so that, for a new input, the head outputs the correct label from a fixed set.

You are coming to this with no prior mastery, so this stands alone here. But in your notes you already have 'Fine-Tuning' as a parent concept, and this is the specific case where the task is label prediction rather than text generation. Contrast with Instruction Tuning: there the output is natural language following a direction; here it is a fixed category. Catastrophic Forgetting is your main risk: if you train on too many classification examples, the model may forget how to write fluent text — a danger to keep in mind even while you are learning the mechanics. The efficient way to do this in practice is with LoRA, where only a small set of extra numbers is trained.