How do students work with datasets in the SAAI Research Program?

The program treats dataset quality as part of the research problem rather than as a hidden preprocessing step. Students learn where data comes from, what each field represents, how labels were produced and whether the dataset is suitable for the proposed question. They inspect missing values, duplicates, class imbalance, inconsistent labels, noisy examples and possible data leakage. Depending on the project, learners may clean an existing dataset, combine multiple datasets, create a small custom dataset or define annotation rules for new examples. They record important decisions so another reviewer can understand how the data was prepared. Students also compare how model performance changes when the dataset, sampling method or label definition changes. This helps them understand that a model’s result is shaped by the quality and assumptions of the evidence used to train it.