The AVS dataset contains derivative results from simultaneous magnetoencephalography, eye-tracking, and structural MRI recordings of five participants performing natural scene viewing and verbal captioning tasks. Participants freely explored 4,080 semantically diverse images selected from the Natural Scenes Dataset, with some trials including verbal scene descriptions. The dataset includes fixation-locked neural recordings, binocular gaze data, anatomical MRI scans, behavioral responses, and task event markers. It is designed to support research on active, naturalistic vision by capturing high-temporal-resolution neural dynamics during self-directed visual exploration, particularly for studying how memory encoding, rather than visual processing demand, influences fixation duration on natural scenes.