Fine-Tuning LLMs for Classification

Building a large language model from nothing involves two very different jobs. The first is writing the code that defines the model and its training procedure. The second is running that training on a large collection of text. Once training is finished, the learned values inside the model, called its weights, capture patterns from that text. Once a model exists, a practical question follows: what can someone do with it, and what can they do without starting over?

The source material makes an encouraging observation. Many pretrained LLMs are available as open source models, and they can serve as general-purpose tools. The important phrase is that they can be used to write, extract, and edit texts that were not part of their training data. A pretrained model is therefore not limited to reproducing the documents it learned from. It can handle new text it has never seen.

A pretrained LLM can be used in two broad ways, and it helps to keep them apart. In the first, the model is used as it is. It acts as a general-purpose tool for tasks such as writing, extracting, and editing text. In the second, the model is adjusted on a specific task using a relatively smaller dataset. The source says this adjustment improves performance on that task while reducing the computational resources needed.

The distinction matters because both paths start from the same pretrained model. The choice depends on the job. If general text work is enough, the model can be used directly. If the goal is strong performance on one particular task, the model is taken through an additional adjustment stage on data for that task.

A Large Language Model that has been pretrained is a general-purpose starting point. It has learned from a broad body of language, but it has not been set up for any one job. Fine-tuning is the step that connects that general model to a specific task. The source describes fine-tuning as the process of adjusting a pre-trained LLM to perform a specific task.

The word that matters in that definition is adjusting. Fine-tuning begins with a model that already exists and changes it so that it performs a particular task. The starting point is general and the ending point is specialized. Fine-tuning continues to adjust the model's weights on data that belongs to the target task, which is why the target task determines which data is chosen.

In the diagram, the architecture itself is the same in both versions. What is different is the set of weight values, which have been adjusted toward the target task. The model does not become a different kind of object. It becomes a version of the same model that is better suited to one job.

The second idea in the source concerns where the starting weights come from. The source explains that after implementing the pretraining code, a learner can reuse openly available model weights and load them into the architecture they implemented. This lets the learner skip the expensive pretraining stage when fine-tuning. Reuse does not mean the weights are discarded or ignored. It means the costly step that produced them does not have to be repeated.

Read the timeline from left to right. The architecture is implemented first. The pretraining stage would normally produce weights, but the source describes reusing openly available weights in its place. Those weights are loaded into the architecture, and the smaller task dataset is then used for fine-tuning. The pretraining stage is absent from this path because the weights were already available.

The phrase predefined categories is the key part of the definition. The classifier does not invent a new label for each input. Someone decides the set of possible categories before the model is used. For a spam filter, the predefined categories are spam and not spam. Every incoming text is mapped onto one of those two options.

The source gives the example of fine-tuning an LLM to identify spam messages. The scenario below follows that goal step by step, using a single incoming message. The scenario is an original illustration written for this article, not a quotation from the source.

The same pipeline applies to every incoming message. A message that does not look like spam is mapped to the other predefined category, not spam. The number of possible outputs is fixed in advance, and that is what makes the output a classification.

A fine-tuned classifier and a general-purpose LLM can both work with text, but they produce different kinds of output. A classifier chooses from a set of fixed options. A general-purpose model used for writing produces free-form text. The source describes classification as producing a choice among fixed categories rather than newly generated text.