The Impact of Training Data Quality

Imagine a translation model that has been carefully built and trained. Its design is sound, and it handles many translation requests well. Then you ask it to translate English into Vietnamese, and it cannot produce Vietnamese at all. The cause is not a flaw in the model's design. The cause is that Vietnamese never appeared in the material the model learned from.

The two properties named in this principle are different, and it helps to keep them apart. Quality concerns whether the data is accurate and useful. Relevance concerns whether the data actually covers the situations the model will face. A training set can be accurate and still fail on relevance, because accurate examples of the wrong things teach the model very little about the work it is actually asked to do.

A useful way to think about this is as territory. The training set defines the territory the model has explored. Anything inside that territory is something the model has had a chance to learn. Anything outside it is effectively unknown, and the model's behavior there is unreliable. Data therefore does more than influence performance in general. It sets the boundary of where performance can be trusted.

Follow the chain step by step. First, a category of examples is missing from the training data. Second, the model never receives examples from that category, so it has no learned patterns for it. Third, a related input arrives that depends on that missing category. Finally, the output is weak. The diagram below lays out this sequence using the translation case.

The same chain explains the image case. An image model trained only on animals has no relevant experience with plants. When a plant photograph arrives, the model meets an input from a category it never covered, and the result is weak performance. In both cases the gap is in the data, and the weakness appears in the model's output.

One way to check a model is to ask which regions of the possible inputs its training data covers, and which regions are empty. Regions with training examples are inside the model's territory. Regions without examples are the places where its behavior cannot be trusted.

Reading the coverage this way turns a vague worry about a model into a concrete question. For any new input, you can ask whether training examples like it exist. If they do not, the model is operating in empty territory, and you should expect its results there to be unreliable.

Two training sets can both be high quality, yet produce very different results on a specific task. What separates them is how closely the data matches that task. Accurate data that does not cover the needed examples leaves the model without the experience it requires. Data that closely matches the task gives the model relevant examples to learn from.

If the training data determines what a model can do, the natural response to a weak skill is to change the data. The course material advises that to improve a model on a certain task, you might include more data for that task in the training set. The reasoning follows from the chain traced earlier. A task that is poorly covered needs more examples of that task, so the model has something to learn from.