Fine-Tuning and Evaluating the Model

Before you fine-tune a model, it helps to know how well it already performs on the data you care about. That baseline gives you something to compare against once training has run. Without a starting measurement, a later improvement or regression has no reference point, so you cannot tell whether the training changed anything meaningful.

In this article we measure a classifier's accuracy on three separate datasets: the training split, the validation split, and the test split. We then look closely at what the resulting numbers tell us at this early stage, while the model has not yet been fine-tuned and its LoRA matrix B still holds its starting values.

Accuracy is the share of predictions the model gets right. For a classifier, each example in a dataset has a true label, and the model produces a predicted label. Accuracy counts how many of those predictions match the true labels and reports that count as a fraction of all the predictions made.

We compute accuracy three times, once for each split, because each split answers a different question. The training split is the data the model learns from, so its accuracy tells you how well the model fits the examples it has seen. The validation split is held apart from the main learning signal and is used to check how the model behaves on data it was not trained on. The test split is kept for the final check. Comparing the three numbers shows whether performance is similar across the splits or whether one split stands out.

Before you look at any output, make a prediction. The model has not been fine-tuned yet, and LoRA matrix B has been initialized with zeros. Think about whether you expect the three accuracy values to be close to one another, and whether you expect them to match the values from chapter 6.

The call pattern above is an illustration of the helper's arguments as described in the course, not a copy of the official listing. Notice what the three calls share: the model and device are identical, and only the loader changes. Keep one question in mind as you read on: what would have to change for the printed numbers to differ from what we expect?

The three datasets play different roles, and each accuracy value tells you something different about the model on its own split. The training split reflects what the model has been fitted to. The validation split reflects behavior on examples held apart from the main learning signal. The test split is reserved for the final check and should not guide decisions made during development.

The diagram shows the flow in plain terms. Each loader feeds the same model, the model's predictions are scored, and the three results are printed separately. Because the same model is used for all three, a difference between the numbers reflects the data each split contains rather than a different model. Here the three values are close, so the model performs about the same on each split.

The reason the initial numbers match chapter 6 lies in how LoRA matrix B starts out. Initializing B with zeros makes the product of matrices A and B a zero matrix. Adding a zero matrix to the original weights does not change those weights. So at the moment of initialization, the LoRA addition has no effect on the model's behavior.

Read the chain from top to bottom. Each link depends on the one before it. The explanation for the matching numbers rests on a single link: the zero starting value of B. If that link were broken, the rest of the chain would no longer guarantee the same results.

Suppose the printed initial values differ from the chapter 6 values. The chain from the previous section tells you where to look. Matrix B is the first link, because a nonzero B makes AB nonzero, which changes the weights and removes the guarantee that the numbers match. The diagram below turns that reasoning into an order of checks.

In words: first check the initialization of B. If B is not zero, the weights have already changed, and that alone explains the mismatch. If B is zero, the next place to look is the AB product and the step that merges it into the weights, since those are the links that should leave the weights unchanged. Only after the weights are confirmed should you look at the data split, because the same model is used for all three measurements and the splits are the remaining source of difference.