Understanding the Dataset and Prompt Styles

Imagine a single training example with three pieces. The first is a request to fix a word. The second is the misspelled word itself. The third is the correct answer. In this article the misspelled word is 'Ocassion' and the corrected form is 'Occasion'. Keep this one example in mind, because the rest of the article follows it through two different layouts.

Before reading on, try to answer this question for yourself: if a model is going to learn from this example, what has to be decided about the three pieces before training can begin? Write down your answer, then compare it with the explanation that follows.

A training example does not reach the model as a neat record with labeled slots. Someone has to write the three pieces out as one block of text, decide where the request goes, where the misspelled word goes, and where the answer goes, and mark those places so the model can tell them apart. The diagram below traces that path from the raw record to the text the model trains on.

Alpaca was one of the early LLMs to publish a detailed account of its instruction fine-tuning process, and its layout is widely used. The Alpaca style opens with a framing sentence and then uses labeled sections for the instruction, the input, and the response. The rest of this chapter uses this layout, so when a training entry appears later, assume it follows this structure unless the text says otherwise.

Here is the spelling example written in the Alpaca style. The framing sentence and the exact label wording shown here are an illustration of the pattern, applied to this example.

Phi-3 comes from Microsoft and is included here to show that prompt styles vary across models. Its style is simpler than Alpaca's. Instead of labeled sections, it uses designated user and assistant tokens: the user token marks the request, and the assistant token marks the reply.

The same spelling example in the Phi-3 style is shown below. The token strings are written in angle brackets here for readability; the exact spelling of the tokens is a property of the model's own format, not something this article defines.

Both layouts encode the same spelling correction. The misspelled word and the corrected form are identical in both. What differs is the scaffolding around them. The diagram lines up the markers position by position, so you can see which marker in one style corresponds to which marker in the other.

Because both styles can encode the same content, a difference in layout is not the same as a difference in meaning. The spelling fix is still the spelling fix in either layout. But the training text itself is different: a model trained on one layout sees one set of markers around the request and the answer, and a model trained on the other sees a different set. The style therefore decides the exact text that the training process works on, even when the underlying pair of request and answer is the same.

The chapter includes an exercise that makes this concrete. You fine-tune a model with the Alpaca layout first, then try the Phi-3 layout from the comparison above, and then observe whether that change affects the quality of the model's responses. The point is to see the effect for yourself rather than assume it. Notice that the exercise asks you to observe, not to predict a fixed outcome.

A training entry tells the model two things: what to receive and what to produce. In the Alpaca layout, the boundary between request and answer is the Response label; everything after that label is the target the model should reproduce. In the Phi-3 layout, the boundary is the assistant token, which opens the reply. Locating these boundaries is the first thing to check when reading any formatted entry, because the target response is defined by where it begins.