Competition & Olympiad Problems
Competition problems that separate reasoning from pattern-matching. Each item ships with a complete solution, a final answer, and — critically — the results of running frontier models on it, so the difficulty is measured, not assumed. Many items are ones today's best models still get wrong.
Coverage
- Olympiad mathematics
- Informatics
- Combinatorics
- Algebra
- Number theory
- Competition STEM
Deliverables
- Problem statements
- Full solutions
- Final answers
- Per-model run records
- Difficulty & knowledge-point tags
From source to acceptance
We don't hand-label a pile and ship it. Every category moves through a closed, instrumented loop — generated to a brief, checked by machines, adjudicated by experts, and traceable end to end — but the path each data type takes is its own.
- 01
Curation & Normalization
Olympiad-grade problems are sourced from competition tradition and normalized into clean, self-contained statements with consistent notation and no ambiguity of intent.
- 02
Worked Solutions
Each problem is paired with a complete worked solution and an exact final answer, giving an unambiguous oracle that automatic grading can settle without human judgement.
- 03
Measured Difficulty
Every item is run repeatedly across a panel of frontier models and the pass counts are recorded; expert curators then adjudicate the borderline cases, so difficulty is empirically measured rather than editorially guessed.
- 04
Knowledge-Point Tagging
Grade level, domain, and the specific techniques each problem demands are tagged, turning raw pass counts into a navigable map of where and why models fail.
- 05
Stratified Sampling
A stratified draw across grades, domains, and measured hardness fixes the final distribution, letting a set target precisely the frontier where a given model breaks.
Every run emits a learning signal that feeds back into the source set — the pipeline tightens itself, batch over batch.
See the data itself
One real, trimmed sample from this category — the scenarios it serves, why it matters for training, and the shape of the data as delivered.
Where it’s used
- Benchmarking mathematical reasoning against a measured difficulty curve
- Mining hard negatives models repeatedly fail
- Training with verifiable final answers as reward signal
Why it matters for training
Essential
A calibrated hardness signal: because each problem carries per-model pass counts, you can target exactly the frontier where a model breaks.
Notable features
Combinatorics
Permutations & combinations, algebraic identities, extremal analysis, case analysis
Reveal answer
8 frontier-model attempts recorded · correct count: 0 / 8 — none solved it.
Need a sample or a custom build?
Tell us your spec and scale — we deliver to order.
← Back to Data