VQA: Verifiable question answer problems with a precise numeric or symbolic final answer.

Subproblems*: Problems structured into intermediate steps leading to a final answer.

Formal Proofs*: Problems formalized in Lean for machine checkable proofs.

Frontier Targeting: Problems calibrated to challenge the most capable models available today.

Configurable pass@N: Target solve rates at specified attempt counts.

Agentic*: Multi step problems designed for workflows that require tool use or multiple API calls.

Computational: Requires code execution or algorithmic search, often with nontrivial runtime.

Analytical: Proof oriented or insight driven problems that reward structural reasoning.

Undergraduate: University level problems and contest style problems.

Graduate: Qualifying exam level problems and advanced coursework level problems.

Post Graduate: Research level problems requiring deep domain expertise.

Mid Training: High quality mathematical content for continued pretraining and fine tuning.

RL Post Training: Reward verifiable problems designed for reinforcement learning pipelines.

Text Only: Pure symbolic and textual problem statements with numeric or symbolic answers.

Multimodal: Problems involving diagrams, figures, or visual reasoning.