Model purpose. The model was trained and employed to extract species observations from videos recorded in shallow-deep (7-105m) underwater hard substrate habitats in the Koster National park at the northwest Swedish west coast, to assess historical changes in vertical distribution and relative abundance of 17 benthic invertebrate taxa. Taxonomic scope. The model was trained to identify the following benthic invertebrate taxa (ranging from species-family level): Protanthea simplex, Sabella pavonina, Geodia barretti, Mycale lingua, Polycarpa pomaria, Serpulidae, Molgula spp., Ascidia spp., Ciona intestinalis, Porania pulvillus, Acesta excavata, Actiniidae, Phakellia ventilabrum, Munida spp., Lithodes maja, Caryophyllia smithii, and Alcyonium digitatum. Model input. The video material consists of archived ROV footage of the area west of the island of Yttre Vattenholmen, stored at the facilities of Tjärnö Marine Laboratory, collected by the ROV pilots Joel White, Tomas Lundälv, and Roger Johannesson. Model outputs: Detections made by the are published as a downloadable Darwin Core formatted GBIF occurrence dataset: https://doi.org/10.15468/rzhmef Temporal scope. The model was trained and used on movies collected in the years 1997-2023. Geographic scope. The model was trained on footage from rocky slope and wall habitats from Koster National Park, at ~10-100 meters depth.