Fine-Tuning GPT-2 for Text Classification

Many learners who load a pre-trained model for the first time make the same quiet assumption. They run the loading call, see no traceback, and conclude that the model is ready to use. The reasoning feels natural: if something had gone wrong, the program would have complained. This section challenges that assumption. Whether the weights are correct is something you check after loading, not something the absence of an error answers for you.

The work in this section happens in a fixed sequence. Each step leaves the GPTModel instance in a different state, and the order matters. First, you call load_weights_into_gpt with the gpt instance and a params object. This places the OpenAI model weights into the GPTModel instance. Second, you move that instance to the target device with gpt.to(device). Third, once the model is loaded, you call the generate function on it, passing the model along with settings such as the maximum number of new tokens and the context size. Finally, the resulting token IDs are converted back into readable text with token_ids_to_text, and the text is printed.

Notice the position of the check. It sits at the end of the chain, after the text has been generated and decoded. The loading call does not hand you a verdict. That is why the source moves directly from the loading step to a generation step. The output is the evidence you use to decide whether loading succeeded.

Before you run the generation step, decide what would count as evidence. Compare two kinds of signal: a program that stops with an error, and a program that finishes and prints text. Only the second one tells you anything about how the model behaves, and even that depends on whether the text is coherent. A program that finishes is not the same as a model that works.

The source makes a specific claim: because the model can produce coherent text, we can be confident that the weights were loaded correctly. The test is about the behavior of the whole system after loading, not about a single line. Loading is a mapping from stored parameters into the structure of a GPTModel instance. If that mapping is right, the instance behaves like the pre-trained model and continues a prompt sensibly. If the mapping is wrong, the output stops being a reliable sign of anything.

The source's own example uses the prompt Every effort moves you. Its recorded output continues that sentence into fluent English about finding a new way to practice something. That is the kind of result the source treats as the sign of a correctly loaded model. You do not need to know the internal layout of the weights to recognize whether the continuation reads as coherent language.

A pre-trained Large Language Model 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 word that matters in the definition is adjusting. Fine-tuning starts from 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. What stays the same is the underlying pre-trained model you began with; what changes is how it is adjusted for the target job.

A classification task produces a choice among fixed categories rather than newly generated text. This contrast is the central idea of this section. A text generator produces free text as its output. A text classifier answers a different question: which of the known categories does this input belong to? Identifying spam messages is one example of text classification, and for a spam filter the predefined categories are spam and not spam.

The source gives the example of fine-tuning an LLM to identify spam messages. The scenario below is generated for illustration and follows that goal with a single incoming message. The diagram traces the message from input to category.