We generate mathematical problems for AI training. Experts design the tasks and check the reasoning.
Synthetic generation, mathematical review and model evaluation.
Built around the capabilities you need to improve.
01 / The work
Training tasks, expert validation and original evaluation sets. Each project starts with a defined capability, a data format and acceptance criteria.
Expert-designed synthetic data
Experts define the reasoning skills and task families. Our generation tool produces candidate problems, and mathematical review checks the tasks and solutions against your brief.
Mathematical quality control
Experts check problem statements, proposed solutions and model outputs. Review identifies gaps, invalid steps and ambiguous assumptions, with explanations a research team can use.
Original expert-level evaluations
Custom mathematical evaluations, including HLE-style short-answer and multiple-choice formats. We define the target skills, prepare original questions and agree on scoring and difficulty calibration.
HLE: Humanity’s Last Exam. The service is original, project-specific evaluation work, not a resale of the public benchmark.
Subject to technical scoping
Discuss formalizing mathematical statements and proofs in Lean 4. The first step is to confirm subject coverage, suitable expertise, dependencies and the proof-checking requirements.
Formal verification checks the encoded statement. Reviewing whether it captures the intended mathematics is a separate part of the specification.
A clear problem, a defensible reference answer and a review method appropriate to the task. A model failing a question is not, by itself, evidence that the question is useful.
We agree on originality checks, difficulty, format, review requirements and revision rules before production. Automated grading is offered only where a reliable checker can be specified.
02 / Our approach
Generation creates candidates. Expert judgment determines which tasks are correct, useful and ready for the agreed purpose.
Define the reasoning skills, subject coverage, constraints and acceptance criteria with your team.
Capability · SpecificationUse the generation tool to produce candidate tasks. Inspect a small batch before expanding production.
Generation · First batchCheck mathematical correctness, worked solutions and ambiguity. Reject or revise tasks that miss the brief.
Review · RevisionEvaluate difficulty against the agreed models and settings. Deliver the accepted tasks with review records.
Evaluation · DeliveryOur view: better generation needs better judgment.
Human-authored tasks remain part of the mix where the brief calls for them.
03 / Deliverables
An illustrative delivery structure. The exact fields and review methods are agreed for your project.
04 / Expertise
A network of 25–30 mathematical experts. Task design, independent problem-solving and careful review are central to the work.
Backgrounds of individual experts in the network.
Led expert teams creating mathematical reasoning data at Mercor and develops synthetic-data pipelines for model evaluation.
05 / Start a conversation
Tell us what your model needs to get better at.
Start with the task, the difficulty and the outcome you need.