A BIG-bench task testing whether a model learns arithmetic operations followed by an unusual +1 rule from examples.
unassessed
| Category | math |
|---|---|
| Subcategory | few-shot arithmetic rule induction |
| Page status | unknown |
| Metric | exact_str_match |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 6000 |
| Dataset licence | Apache-2.0 |
| Publisher | Google BIG-bench |
Modified Arithmetic gives two numbers and five worked examples using an operation. The model must complete a sixth example. In the challenge subtasks the ordinary operation is followed by adding one, while control subtasks omit the extra one.
Free-text numerical completion across six subtasks: three-digit addition, three-digit subtraction, and two-digit multiplication, each with and without +1.
No model card in ModelSpec reports this benchmark yet.