When a model is trained, the dataset provides examples and answers, but the rows of the dataset are not what get updated over and over. The dataset acts as the teaching material. The thing that changes is the model's internal parameter values, often called weights.
This distinction sounds technical, but it is the key to understanding learning. If a learner thinks the model improves by storing the raw data directly, it becomes harder to understand generalization, overfitting, and evaluation later on. Training is about adjusting internal settings so the model behaves differently on inputs, not about rewriting the facts in the table.
In a simple supervised setup, the model sees an example, produces a prediction, compares that prediction to the known target, and then updates its parameters so similar mistakes are less likely in the future. The data stays the same. The model changes.
That is why people talk about fitting a model to data. The data is the reference point. The model is the flexible object being nudged and shaped in response to what the data shows.