A 30-item BIG-bench JSON template that splits one- to three-digit addition into three digit-length subtasks, as an example of nested JSON tasks.
unassessed
| Category | math |
|---|---|
| Subcategory | BIG-bench nested JSON template; addition split by digit length |
| Page status | unknown |
| Metric | exact_str_match (per subtask); normalized_aggregate_score (parent) |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 30 |
| Dataset licence | Apache-2.0 |
| Publisher | Google (BIG-bench organizers) |
simple_arithmetic_json_subtasks asks a model to complete addition equations, then groups those equations into one_digit, two_digit and three_digit directories, ten items each. The parent task.json holds only metadata. Some two_digit and three_digit lines mix addend widths. The organizers wrote it to show how BIG-bench nests JSON subtasks, not to measure arithmetic. It is tagged "example task". The thirty equations are the same pool as simple_arithmetic_json, rearranged into three files. English-script, text-only, zero-shot.
Free-text numeric completion of an addition equation, reported both per digit-length subtask and as a BIG-bench normalized aggregate across the three subtasks.
No model card in ModelSpec reports this benchmark yet.