Scale AI and Hugging Face have released FineART, an open bimanual robot dataset for researchers and developers training multi-step manipulation policies. It holds 40,543 teleoperated episodes, 1,718 hours across 151 tasks, recorded on a Trossen Stationary AI robot, and every episode carries human-written subtask labels, 533,913 in total. About 77 percent of the hours are pick-and-place and sorting, with the rest covering contact-rich work such as textiles, cables, tool use, and precision insertion. Open robot demonstrations are often labeled with only a task-level instruction, so the intermediate steps are not available as training targets. FineART-VLA, which Scale and Hugging Face describe as the first open System 1-System 2 model in LeRobot, predicts its next subtask in language and then turns that plan into continuous actions inside one policy. Mid-training on the annotations raised success on spatial instructions from 32 percent to 100 percent. With human subtask guidance, success on unseen long-horizon tasks rose from 16 percent to 76 percent. On a different robot, the policy matched baseline performance with one-tenth the fine-tuning data and generalized zero-shot to tasks it had not seen. The dataset, weights, and training code are on Hugging Face under a CC BY 4.0 license, and Scale reports the real-world evaluation covered 3,400 rollouts.