Labelled 30-second Acoustic Clips of ALOHA Cabled Observatory

This dataset contains the labelled 30-second hydrophone audio clips used to train and evaluate ACO-D, an attention-based multiple instance learning (MIL) detector for signal-agnostic transient detection in long-term passive acoustic monitoring. The clips are derived from ALOHA Cabled Observatory (ACO) recordings (mono, 4 kHz) and are annotated into three categories: stationary (background, including vessel tonal noise), transient (clear non-stationary events), and subtle_transient (weak events, used for evaluation only). The set comprises 3,655 stationary, 1,062 transient, and 3,475 subtle_transient clips; stationary and transient are split train/validation/test in an 8:1:1 ratio. The accompanying code is available on GitHub. See README.md for the full directory layout, file-naming convention, and usage.