AA-LCR-derived
RL environmentLong-context reasoning tasks with structured JSON verification.
Ulam
Research, reasoning and software tasks derived from named benchmark formats.
Environments & tasks
17 listed items17 items
AA-Omni-derived
RL environmentTasks testing factual reliability, calibration and abstention.
CritPt-derived
RL environmentPhysics tasks with executable checks; private environment delivery.
CyberSec-derived
RL environmentDefensive repair tasks with partial-credit verification.
GDP.pdf-derived
RL environmentLong-horizon document tasks requiring evidence-backed decisions.
GDPVal-derived
RL environmentProfessional-work tasks with multi-part rewards; Harbor-compatible delivery.
GPQA Diamond-derived
RL environmentGraduate-science tasks with verifier code and repeatable checks.
Humanity's Last Exam-derived
RL environmentFrontier reasoning tasks with executable verification.
Math RL
RL environmentFoundational research-oriented mathematics RL suites with task-specific exact, structured-output or certificate verifiers. Public samples are available; private suite terms require inquiry.
MathH RL
RL environmentMedium-to-hard mathematics RL suites using exact and certificate verifiers, including theorem-tactic and structured-witness checks. Public samples share the Math-RL-Tasks release.
MathR RL
RL environmentFrontier research-math RL suites for agent training and certificate generation, with task-specific exact, witness and construction checks. Public samples share the Math-RL-Tasks release.
SciCode-derived
RL environmentScientific coding tasks with weighted numerical checks.
SciDiamond-derived
RL environmentAdvanced-science tasks with deterministic answer verification.
Tau-Banking-derived
RL environmentBanking tool-use tasks graded on state and policy compliance.
Terminal-Bench Science Hard
RL environmentDifficult scientific tasks with hidden numerical evaluations.
Terminal-Bench-derived
RL environmentTerminal-agent tasks with executable tests and structured checks.
Terminal-Bench-Science-derived
RL environmentScientific-agent tasks using checks, certificates and stateful scoring.
These are supplier descriptions, not quality certifications. Confirm samples, licensing, pricing and current availability directly.
Public benchmarks & research
Public research is separate from commercial inventory. Check original licenses for training rights.
ErdősBench
Public benchmarkResearch-math benchmark with public smoke tests and private evaluation.
SimoBench
Public benchmarkOlympiad-proof benchmark with graded reasoning; not formal proof certification.
Sources & verification scope · 4
Cross-checked against official sites, catalogues and public profiles. Headcount refers to the company at the source date, not its expert network or RL team.
- Ulam Math-RL-Tasks official public dataset cardofficial · 2026-09-18
- Ulam 공개 자료company_profile · 2026-09-17
- Ulam 공개 자료official · 2026-09-17
- Ulam 공개 자료official · 2026-09-18