Crystal Math Preview: 1,000-10,000 Olympiad and Competition Math Problems

Crystal Math Preview is a collection of 1,000 to 10,000 mathematical reasoning problems released by ycchen in February 2026 to accompany a research preprint. The dataset focuses on olympiad and competition-level mathematics, featuring specialized configurations derived from high-reasoning budget rollouts. It serves as an early-access version of a larger planned release for training and evaluating mathematical reasoning models.

Use Cases

• Fine-tuning language models for mathematical reasoning using the high_pass_rate field
• Evaluating model performance on olympiad-level competition math
• Analyzing reasoning paths in high-budget rollout scenarios for RLVR

Strengths

• Contains 1,000 to 10,000 competition-level math problems
• Includes high_pass_rate metadata from 5x rollout reasoning budgets
• Specifically tagged for Reinforcement Learning from Verifiers (RLVR) tasks

Limitations

• Explicit contamination warning: includes problems drawn from common evaluation benchmarks
• Small sample size as a preview release compared to the planned official version
• Limited documentation on specific source origins for individual problems

Provenance

Source

Competition math sources and olympiad benchmarks

Collection Method

Sourced from existing benchmarks and augmented with high-reasoning budget rollouts

Freshness

Last updated February 2026.

Geography

Global (English language)