A 250,000-item BIG-bench task that counts intersections among templated circles, polygons and segments, scored as 41-way multiple_choice_grade.
unassessed
| Category | math |
|---|---|
| Subcategory | count intersection points among 2–6 templated shapes (BIG-bench JSON) |
| Page status | unknown |
| Metric | multiple_choice_grade |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 250000 |
| Dataset licence | Apache-2.0 |
| Publisher | Google (BIG-bench collaboration) |
intersect_geometry (README title: Intersection Points) gives English template sentences that place line segments, polygons or circles in the plane, then asks how many intersection points exist. Choices are the integers 0 through 40. Authors Paul Vicol and Jascha Sohl-Dickstein generated the items to test geometric counting, not fluent reading. Five subtasks use 2, 3, 4, 5 or 6 shapes. It is not [geometric_shapes](geometric_shapes.md), which names an SVG path, and it is not in [BIG-Bench Hard](bbh.md).
41-option multiple choice (0–40) with preferred metric multiple_choice_grade. append_choices_to_input is false. Canary GUID embedded. Root task.json holds metadata only; examples live in shapes_{2–6}/task.json. Dummy-model header reports 249,999 multiple-choice queries.
No model card in ModelSpec reports this benchmark yet.