Loading Pretrained Models

Suppose you want a language model to decide whether an incoming email is spam. The obvious question is where the model should start. Should you build a new model from nothing and teach it to recognize spam directly, or should you begin with a model that already knows how to work with language? The source you are studying describes the second route, and this article explains why it works and what loading a pretrained model actually does.

The core difficulty is that a model which was only pretrained has learned one habit very well: it continues text, one token at a time. Its natural behavior is to predict what comes next in a passage. A spam classifier needs something different, namely a single answer chosen from a fixed set of categories. Before reading on, try to decide what a pretrained-only model would output if you handed it an email.

The difference between the two stages lies in the expected output. During pretraining, the model is trained to predict the next token over its vocabulary. During classification fine-tuning, the model is trained to output a correct class label. The fine-tuning stage does not reinvent how the model reads its input. It changes the target the model is pushed toward.

Read the two columns together. The left column shows the habit the pretrained model already has. The right column shows the answer the classifier must give. The input column is identical in both, which is the point the next section develops.

The input to the model during classification fine-tuning is text that has been converted into token IDs. This is the same kind of input the model received during pretraining. Because the form already matches, fine-tuning does not require a new input format. What it adds is a different expected output: a class label rather than the next token.

Follow a single example from start to finish. The email text is first converted into token IDs. Those IDs are passed into the model. Instead of producing a continuation, the model is expected to produce one class label from its fixed set of categories. The diagram below lays out this path for the spam case.

Classification fine-tuning does not begin from nothing. Before fine-tuning an LLM, the pretrained model is loaded as a base model. The base model is the starting point, and the fine-tuning process then adapts it toward the classification goal. Loading is therefore the first step in the sequence, coming before any input is passed in and before any class label is expected as output.

The key idea is the order of operations. The model is loaded first, and fine-tuning is applied to that already-loaded model. Loading is not the fine-tuning itself. It is the step that gives fine-tuning something to work on.

The model loaded in this workflow is a pretrained GPT-2 medium model with 355 million parameters. Instruction fine-tuning is the process of adapting a model so that it responds to specific instructions or tasks, and it needs a starting model before that adaptation can begin. In this part of the workflow, the starting model is the pretrained GPT-2 medium model, loaded to serve as the starting point for instruction fine-tuning.

Follow the state across these two stages. At the first stage, the pretrained GPT-2 medium model is loaded. Its state at that moment is the pretrained one: it has 355 million parameters and has not yet been adapted to any instruction. At the second stage, instruction fine-tuning begins from that loaded model. The state carried forward is the same starting point, and the fine-tuning process is what changes the model from there.