Supervised machine learning models depend on training datasets with positive and negative examples. Therefore, dataset composition directly impacts model performance and bias. Given the importance of machine learning for immunotherapeutic design, we examined how different negative class definitions affect model generalization and rule discovery for antibody-antigen binding. Using synthetic structure-based binding data, we evaluated models trained with various definitions of negative sets. Our findings reveal that high out-of-distribution performance can be achieved when the negative dataset contains more similar samples to the positive dataset despite a lower within-distribution performance. Furthermore, leveraging ground truth information, we show that binding rules discovered as associated with positive data change based on the negative data used. Validation on experimental data supported simulation-based observations. This work underscores the role of dataset composition, including negative data selection, in creating robust, generalizable, and biology-aware sequence-based ML models.