Initializing and Preparing the GPT Model for Fine-Tuning

A pre-trained GPT model has already learned general patterns from a large amount of text. That learning is valuable, but the model was not built to answer one specific question, such as whether a message is spam or not spam. Before it can be fine-tuned for a classification task, its final layer has to be adapted so that it produces one output for each class the task requires.

This article follows the setup step that does exactly that, and then prepares the model to run on an available device. The setup is short, but every line changes something specific. As you read, ask two questions about each line: what object is being changed, and what value does it end up holding? Keeping those two questions in mind will make the rest of the article much easier to follow.

Think of the pre-trained model as a stack of components. Most of the stack is the GPT architecture itself, which turns input text into internal representations. The last component is the output layer, which turns those internal representations into the outputs that the model was originally built to produce. For a language model, those outputs are scores for possible next tokens, which is why the original output layer is sized to the model's vocabulary.

A classifier needs something different. It must produce one output per class, and the number of outputs should match the number of categories in the task. The setup therefore leaves the body of the model alone and swaps only the final layer. The diagram below shows the change: the input width stays the same, while the output side is redrawn for the classes.

The replacement is a single assignment to the model's out_head attribute. The right side of the assignment creates a new torch.nn.Linear object. The left side stores that object in place of the old layer. Nothing else in the setup code redefines the model, so the rest of the GPT architecture stays exactly as it was loaded. Only the final layer that turns internal representations into outputs is changed.

Read the anatomy from left to right. The attribute on the left is the location being overwritten. The call on the right defines the new layer's shape. The in_features value is 768 in this setup, and the out_features value comes from num_classes. Changing the number of classes therefore changes only the output side of the layer, while the input side stays fixed at 768.

The setup can be read as a sequence of state changes. Each line leaves the model in a new state, and the next line builds on it. The diagram traces the model from the pre-trained version through the output layer replacement and the device placement. The states are listed in the order the code runs.

Notice what does not appear in this trace. The body of the GPT architecture is never reassigned, so its learned parameters are carried forward unchanged from the pre-trained model. The only new object is the output layer, and the only new placement is the device the model is moved to. Both the output replacement and the device selection happen before any fine-tuning begins.

After the output layer is replaced, the setup prepares the device the model will run on. The setup code first creates a device object. It selects CUDA when torch.cuda.is_available() reports that CUDA is available, and selects the CPU otherwise. The device is therefore chosen based on what the machine offers, rather than being hard-coded into the program.

The order of the two steps matters for reading the code. The model is first adapted to the classification task by replacing its output layer, and then it is placed on the device where it will run. Reversing the order would not change which layer is replaced, but the setup as written adapts the model first and places it second.