Data and code for a YOLO-based computer vision workflow for minute-level monitoring of oyster valve behaviour

This record contains the quality-controlled development dataset, model artefacts, analysis-ready behavioural records, ancillary heart-rate data, operator-reviewed identity maps, and code accompanying a YOLO-based workflow for automated minute-level monitoring of oyster valve behaviour from top-view infrared video. The development dataset contains 1,396 images and 11,207 open/closed oyster bounding-box annotations, divided into 1,116 training images and 280 held-out validation images. The deposited checkpoint, training outputs and code support checkpoint re-evaluation, longitudinal inference, minute-level aggregation, statistical analysis and figure generation. The behavioural files cover 17 analysed oysters monitored over 19 days. The source records used for the leakage-controlled validation assessments and the approximately 1.5-TB full raw-video archive are not included in the public deposit and are available from the corresponding author upon reasonable request. The deployed checkpoint was trained on an earlier annotation snapshot of the same 1,396 images. A later quality-control pass removed 570 boxes and produced the archived 11,207-annotation dataset. Re-evaluation of the checkpoint against the archived validation labels retained mAP@0.5 = 98.4%. This provenance is documented beside the model artefacts.