Preparing Text Data for Language Models

When a team decides to adapt a large language model, the first question is rarely which training recipe to use. The more useful question is what the adapted model must do. Some projects need one model that can take many different kinds of requests and respond sensibly to each. Other projects need a model that sorts incoming text into a fixed set of labels and nothing else. The two fine-tuning approaches covered here are built around those two goals, and the choice between them shapes the data you collect, the hardware you need, and the evaluation you run.

Instruction fine-tuning aims at flexibility and interaction quality across many complex user instructions. It is the more versatile of the two approaches, but it demands larger datasets and greater computational resources in order to produce a model that is proficient across many tasks.

Classification fine-tuning asks less of both data and compute. It suits precise categorization into predefined classes, such as sentiment analysis or spam detection. The cost of that economy is scope: a classification model's use is confined to the specific classes it was trained on. A new category it has never seen is not something it can handle by design.

Once classification fine-tuning has been chosen, the work proceeds in three stages. Stage 1 is dataset preparation. Stage 2 is model setup. Stage 3 covers fine-tuning the model and evaluating it. Each stage hands its results to the next, so the order matters.

Stage 2 turns the prepared data toward an actual model. Its steps are to initialize the model, load pretrained weights, modify the model so it suits fine-tuning, and implement evaluation utilities. Stage 3 then runs the fine-tuning itself, evaluates the fine-tuned model, and finally uses the model on new data.

The worked example used throughout this topic is a classification fine-tuning project on the GPT model implemented and pretrained earlier. The task is to modify that model and fine-tune it for classification, using a text message dataset that contains both spam and non-spam messages. The text message setting makes the goal intuitive: each message needs one of two labels.

Dataset preparation begins with downloading the dataset. After that comes inspecting and preprocessing it, and only then creating data loaders. Follow one message through that sequence. It starts as raw text with a label attached: a short message marked as spam, for example, or a short message marked as non-spam. Inspection tells you what the rows look like, and preprocessing converts each message into the form the model can read. The data loaders then feed batches of those rows to the model during training.

Messages rarely have the same length, and a model consumes batches of rows that line up position by position. A common way to handle this is to convert each message into a sequence of token IDs and then pad shorter sequences with a filler ID so every row in the batch has the same length. The illustration below uses made-up token IDs to show the idea. It is a generated example of the technique, not a value taken from the source.