Post-training data for frontier AI
Advancing AI through human expertise.
RL environments, evaluations and expert data for frontier AI teams.
What we build
Post-training data and expert teams.
RL environments and verifiers
Verifiable tasks in resettable sandboxes, with graders tested against hacks.
Evals and benchmarks
Private held-out eval sets with pass rates and failure analysis.
Expert data
Demonstrations, preference rankings, rubric scores and gold trajectories.
Managed expert teams
A matched team of engineers, run by our leads and our QA.
Domains
Where we work.
Software engineering
Repository-level fixes and features, judged by hidden tests.
DetailsAgentic workflows
Resettable apps where agents use real tools.
DetailsCybersecurity
Find-and-fix work on real vulnerabilities.
DetailsPhysical AI
Robotics and embodied AI datasets on the same pipeline.
DetailsAnd more
Expert Community
Expertise across domains.
Domain experts working on the next generation of AI.
Our software engineering community.
One of the specialist communities behind our work.
Assessed and interviewed engineers from NxtWave's ~700K-developer community in India, matched to each task by skill level.
- engineers in NxtWave's community
- ~700Kengineers in NxtWave's community
- principal and staff engineers
- ~46Kprincipal and staff engineers
Quality
Checked before delivery.
- Reference solutionEvery task ships with a reference solution that passes its own verifier.
- Verifier checksGraders are checked against correct solutions and invalid outputs.
- Red-teamedRed-teamers try to game every task and grader before hand-off.
- Open-weight runsOpen-weight models are run on every task; pass rates are reported and failure traces reviewed.
legacy-batch-reconcile
ILLUSTRATIVEFix the nightly reconciliation report.
| Attempt | Result | Reward |
|---|---|---|
| Empty solution | FAIL | 0.0 |
| Hack: delete refund rows | BLOCKED | 0.0 |
| Hack: hand-write the report | BLOCKED | 0.0 |
| Reference solution | PASS | 1.0 |