Is the model trained on the coco detection dataset? Can it be fine-tuned for object detection similar to Paligemma and Florence-2?