This dataset contains electroencephalography (EEG) recordings collected for five-class imagined alphabet decoding using an OpenBCI system. EEG data were recorded from 20 healthy adult participants during an eyes-closed alphabet-imagery task. The five target alphabet classes are A, K, N, P, and Z. Each participant completed one recording session for each alphabet class, and each session contained 25 repeated trials. The dataset includes raw OpenBCI recordings, externally generated event-log files, labelled training-segment files, metadata tables, preprocessed MATLAB epoch files, channel-quality audit files, generated figures, and baseline deep-learning result files. The raw recordings were acquired with a 16-channel OpenBCI Cyton Serial Daisy configuration at 125 Hz. The raw channel montage includes Fp1, Fp2, C3, C4, P7, P8, O1, O2, F7, F8, F3, F4, T7, T8, P3, and P4. During preprocessing, channel-quality assessment identified T8 as problematic; therefore, the final exported preprocessed epoch set contains 15 retained EEG channels. The preprocessed release contains 100 alphabet sessions, 2,500 task trials, and 390,000 exported EEG windows. Each 10-second task segment was divided into 156 non-overlapping windows of 8 samples. At the 125 Hz sampling rate, each 8-sample window corresponds to 64 ms. The exported EEG tensor is stored as windows × time points × channels, with a final shape of 390,000 × 8 × 15. The dataset is accompanied by metadata files describing participant-level anonymized information, channel configuration, and task timing. It also includes code and outputs for baseline deep-learning experiments using leave-one-subject-out and calibration-based evaluation settings. The baseline workflow evaluates CNN-LSTM and LSTM-based architectures for imagined alphabet classification. This dataset is intended for research on EEG-based brain-computer interfaces, imagined speech or imagined character decoding, subject-independent EEG classification, calibration-based learning, low-cost OpenBCI acquisition, EEG preprocessing reproducibility, and benchmark development for deep-learning methods.