Dataset containing the raw data collected from individual nanoparticle on mirror geometries (using 80nm nanoparticles and 4-biphenyl-thiol as spacer). This is the data used for the associated machine learning algorithms. To this end scans (each containing 1000spectra) were labelled whether they contain a "picocavity" as True or False. This dataset is then partitioned and used to train the salient feature extraction method to isolate single molecule SERS signals. After this the Siamese CNN was trained to extract wandering correlations as reported in the associated publication, showing how molecules and metals interact on a single molecule level.