Evaluating Finetuned Model Performance

A common assumption among people new to finetuning goes like this: the model has been trained, so it should understand any reasonable request. If an instruction is clear to a human reader, it should be clear to the model. This belief feels natural, and it leads many learners to treat prompt wording as a minor detail once training is finished. This article shows why that assumption can break down. Once a model has been finetuned, the way you phrase and format a prompt becomes a far more important choice than it first appears.

Imagine you finetuned a model on examples where every input looked like the word burger followed by an arrow, written as burger -->. Later you want the model to respond to the burger item, and you consider three prompts: burger, Item: burger -->, and burger --> with a trailing space. Before reading on, decide which of these looks most like the training data and which ones you suspect could cause trouble. Try to say why in one sentence. Your guess does not need to be right. The point is to notice which details you are paying attention to, because the rest of this article depends on exactly those details.

The central principle is that the prompts you send to a finetuned model should match the format of the data used to finetune it. Finetuning teaches the model from examples, and those examples carry a particular shape: a particular arrangement of words, punctuation, and spacing. When you later send a prompt that departs from that shape, you are asking the model to operate on input it was not trained to expect. The source does not claim that every deviation will fail, but it does warn that near-miss prompts can cause issues even when their meaning is clear to a human.

Return to the burger scenario. The training inputs were all written as burger -->. Each of the three candidate prompts departs from that exact string in a different way, and it helps to trace each one separately rather than judging them by overall meaning, since all three are clearly about a burger to a human reader.

Finetuning offers a second benefit that runs in the opposite direction from careful formatting. Once a model has learned the behavior you want from its training examples, you can often send a much shorter prompt than you would need for a base model that has not been finetuned. The source describes the finetuned prompt as much shorter than the prompt used with the base model.

The diagram shows the idea in outline. Before finetuning, the base model needs the instructions and examples inside every request. After finetuning, the behavior lives in the model's training, so the request itself can be short. The exact reduction depends on the task and has to be measured rather than assumed.

The same reasoning explains why the format of the finetuning data itself matters. The source notes that different finetuning data formats can change how well the resulting model performs. The format you choose when preparing data is therefore not only a bookkeeping decision. It shapes what the model learns and, by extension, what kind of prompts it responds to well.

The source does not name one best format for every situation. Instead it recommends experiments. In practice this means preparing examples in more than one format, finetuning a model on each version, and comparing the results on the task you care about. A format that looks tidy on paper may not produce the best behavior once a model is trained on it, and only testing can reveal that difference.