OpenSpatialLogic
# Dataset Card for OpenSpatialLogic ## Dataset Summary OpenSpatialLogic is a handcrafted dataset of 50 riddles which test understanding of spatial relationships in reality. These include questions about compass directions, ordering of bricks within towers after transformations, and permeability of objects in certain configurations. The only way for a model to get better at something is to train on data about it. Large Language Models are bad at understanding spatial reality because their knowledge about the universe is based on relationships between words in a text corpus, not understanding of the universe directly. This data has been crafted so models can be trained to understand these physical relationships better. ## Dataset Structure OpenSpatialLogic contains a question under the "Context" field, and then a "briefResponse" and "detailedResponse" field which both answer the question in varying levels of brevity. ### Example: Context: A tower is built from bottom to top with bricks colored red, green, purple, yellow, then blue. If the tower is turned upside down, which brick is now directly above the yellow? briefResponse: The purple brick will be above the yellow brick. detailedResponse: The tower is turned upside down, which inverts the order of the bricks. The purple brick was below the yellow brick in the previous configuration, so the purple brick is now above the yellow brick.