Full-dataset and commercial access
AIME++ is built for model teams that need large volumes of difficult mathematical reasoning problems with a transparent, deterministic success condition. Each problem resolves to an integer from 0 through 999, enabling exact-match evaluation and RLVR rewards without a learned judge.
This repository is a 157-problem inspection and integration sample released under the MIT License. Teams may use it to validate their data pipeline, run initial model experiments, test exact-match rewards, and inspect the progression from competition-style to research-level mathematics.
Available scale
Ulam AI's broader database contains more than 100,000 problems. Available AIME++ collections include:
| Collection | Available volume | Description |
|---|---|---|
| AIME family | 24,700+ problems | Standard AIME and AIME Hard problems; graduate and research-level problems are excluded from this count |
| AIME-Graduate | 1,000+ problems | Graduate-level problems with the same 0–999 answer contract |
| AIME-Researcher | 100+ problems | Research-level problems with the same 0–999 answer contract |
The public sample represents approximately 132 AIME-family examples, 20 graduate examples, and 5 researcher examples. It is large enough for technical diligence while preserving the overwhelming majority of the catalog for licensed delivery.
Why teams buy the larger collection
- Production scale: tens of thousands of consistent prompt-and-answer pairs rather than a small benchmark.
- Deterministic rewards: every item supports inexpensive, reproducible exact-match scoring.
- Difficulty coverage: the same output contract spans AIME, AIME Hard, AIME-Graduate, and AIME-Researcher tasks.
- Low integration cost: JSONL, stable IDs, explicit tiers, a JSON Schema, validation tooling, and a reference scorer are already defined in the sample.
- Internal authorship: the problems were created internally by Ulam AI, which holds the dataset rights.
- Flexible deployment: collections can support supervised fine-tuning, RLVR, capability evaluation, regression suites, and custom tier mixes under the applicable commercial terms.
The dataset is intentionally answer-only. Each supplied final answer is the authoritative golden solution for its problem. Worked derivations and chain-of-thought traces are not part of the product format.
Commercial delivery can be scoped by
- collection and number of problems;
- AIME versus AIME Hard composition within the AIME family;
- AIME-Graduate and AIME-Researcher volume;
- training, fine-tuning, distillation, RLVR, or internal-evaluation use;
- public-answer versus separated or sealed answer-key delivery;
- custom topic or difficulty targeting;
- JSONL, Parquet, or buyer-specific export format;
- versioning, validation reports, updates, and support; and
- private evaluation execution and scorecards.
Information to include in an inquiry
To make scoping fast, prospective partners should provide:
- intended use;
- desired problem count and tier mix;
- whether the answer key should be delivered, separated, or retained by Ulam;
- model and tool-access constraints;
- required delivery format and security controls;
- desired evaluation or update cadence; and
- target date and license term.
Visit ulam.ai to contact Ulam AI. Final inventory, deliverables, usage rights, pricing, representations, and support are defined in the applicable written agreement.