On segmentation masks, this dataset follows the COCO (Common Objects in Context) annotation standard in. Each annotation includes polygonal segmentation masks representing individual fish instances with each mask manually assessed for quality, represented as class label. On bounding boxes, the dataset follows YOLO annotation standard with individual fish corresponding to a row with format <class_id> <x_center> <y_center> <width> <height>. The dataset consists of 88 images, collected from two different commercial Nordic RAS farms using cameras installed above the surface. The masks have been produced using Segment Anything Model (Kirillov et al., 2023). Two different prompting methods were used, automatic grid prompt and manually annotated bounding box center points. Both sets of masks have been manually assessed for quality. More detailed structure of the annotation files (.json) and description of the data can be found in README.md.