Math: School-Level Question and Answer Pairs from DeepMind

DeepMind released this modular collection of school-level mathematical question and answer pairs in 2019 to evaluate algebraic reasoning in neural models. The data is programmatically generated and organized into specific subsets such as arithmetic__mul to test targeted logic skills.

Use Cases

• Benchmarking model accuracy on arithmetic multiplication using the arithmetic__mul module
• Training sequence-to-sequence models using question and answer pairs
• Evaluating algebraic reasoning through school-level mathematical questions

Strengths

• Authored by DeepMind researchers
• Provides modular subsets like arithmetic__mul for granular evaluation
• Designed specifically to test algebraic reasoning skills

Limitations

• Synthetic generation may lack the linguistic variety of human-written math problems
• Restricted to school-level difficulty levels
• Metadata does not provide specific row counts or column headers

Provenance

Source

DeepMind (Saxton, Grefenstette, Hill, Kohli)

Collection Method

synthetic

Freshness

Last updated January 2024; based on 2019 research.