This dataset is used in the experiments of this publication: K. Vo, M. Vishwanath, R. Srinivasan, N. Dutt, and H. Cao. “Composing Graphical Models with Generative Adversarial Networks for EEG Signal Modeling”, in Proc. of IEEE ICASSP, 2022. The 23-channel interictal EEG recordings from the CHB-MIT epilepsy dataset are used for the experiments. The dataset consists of scalp EEG from pediatric subjects with intractable seizures. We select a subset of 6 patients (chb01-03, chb05-06, chb10) having the same measurement setup, including males and females, 1.5-14 years old. Interictal periods are extracted at least 4-hour away before a seizure onset and after the seizure ends. The signals are low-pass filtered with a cut-off frequency at 50 Hz and scaled to the range [-1,1]. Overall, the dataset contains 43593 signals, from which 70% are used for training and validation, and the other 30% are used as the test set. Each signal is 10-second long (T=10), at a sampling rate of 256 Hz. Additionally, 339 ictal EEG signals are extracted for evaluating epilepsy seizure detection performance.