We collected a dataset of video recordings from 16 male laboratory rats housed in operant conditioning chambers. Each rat was recorded in nine separate sessions over a period of six weeks, yielding a total of 144 videos. From each video, 10,000 frames were extracted, resulting in a dataset of 1,440,000 images. Recording conditions varied across sessions, as rats were assigned to different chambers and recorded with different cameras. As a result, illumination, camera focus, field of view, and other imaging conditions changed across videos. This variability was preserved to assess whether identification models could generalise across recording sessions rather than relying on video-specific visual cues. For each frame, the position of the rat was automatically estimated and a square image region centred on the animal was extracted. All images were then resized to a resolution of 128x128 pixels.