DeepScaleR: 40,000 Competitive Math Problems and LaTeX Solutions

DeepScaleR Preview Dataset contains approximately 40,000 unique mathematics problem-answer pairs compiled by agentica-org in early 2025. The collection aggregates competitive math problems from the American Invitational Mathematics Examination (1984-2023), American Mathematics Competitions, Omni-MATH, and the Still dataset.

Use Cases

• Fine-tuning large language models for mathematical reasoning using the 'problem' and 'solution' columns
• Evaluating LaTeX parsing and rendering accuracy for complex mathematical formulas
• Benchmarking model performance against historical AIME and AMC competition standards

Strengths

• 40,000 unique problem-answer pairs
• Spans nearly 40 years of AIME competition data (1984-2023)
• Standardized LaTeX formatting for mathematical notation

Limitations

• Domain-specific focus restricted to mathematics
• Official solutions may lack the intermediate chain-of-thought steps required for certain reasoning models
• Potential data leakage risk for models previously trained on AIME/AMC benchmarks

Provenance

Source

Aggregated from AIME, AMC, Omni-MATH, and Still datasets by agentica-org

Collection Method

Compiled from historical competition archives and existing math datasets

Time Range

1984-2023

Freshness

Last updated February 2025.