Modifying Pretrained LLMs for Classification Tasks

When a pretrained language model is adapted for classification, one of the first design decisions concerns the output layer. The output layer produces the final result the model returns for each input. For a classifier, that result is a choice among a fixed set of categories, so the design question becomes: how many output nodes should the layer contain? The answer depends on how many categories the task has, and the choice has consequences for how the model is trained.

This article follows one question through a series of small steps. Start with a concrete task, watch the output layer take shape, and then see what changes when the task has only two categories. Keep the question in mind as you read: what determines the size of the output layer?

Take the source's own example. A model must classify news articles as Technology, Sports, or Politics. The task has three classes, so the output layer gets three nodes. Each article passes through the model, and the three output nodes together represent the model's answer across those three categories.

Read the diagram from left to right. Each node is a dedicated slot, and each slot corresponds to exactly one category. The layer has as many slots as there are possible answers, so there are three slots because there are three answers. The labels on the slots are an illustrative assignment made for this example; what matters is that every category has its own output.

The rule is simple to state, but it is worth seeing why it is the default. When the layer has one node per class, every category has its own dedicated output, and the model's answer is read out in terms of those categories. Nothing about the answer has to be inferred from a shared output or a clever encoding. Each category is represented directly.

Put another way, the output layer is sized by the task rather than chosen freely. A task with two categories calls for two nodes, a task with four categories calls for four, and so on. The source describes this as general because it does not depend on any special case and extends to any number of classes.

A binary task has only two possible answers, which raises an obvious shortcut. Instead of giving each class its own node, a single output node could in principle carry the whole answer. The source confirms that this is technically possible for binary classification. Its catch is that using one node requires modifying the loss function, which is the quantity the model minimizes during training.

In the two-node layout on the left, each class has its own dedicated output, and the answer is read out class by class. In the single-node layout on the right, one output is responsible for the whole binary answer. The labels Spam and Not spam are an illustrative binary task invented for this article; the source does not name a specific binary example.

A mismatch happens when the output layer's size does not match the classification task. The check is simple: count the classes in the task and compare that count with the number of output nodes. If the two numbers differ, and the model is not deliberately using the single-node binary shortcut, the output layer is the wrong size for the task.

The last row of the table shows a case that is easy to misread. A single node for a binary task is not a size error. The source says it is technically possible, so the mismatch is about the loss function rather than the number of nodes. Learners should treat a one-node binary layer as a deliberate design choice that requires a matching loss modification, not as a sign that something has gone wrong with the layer.