Off-the-shelf data
Expert-designed tasks, trajectories, and RL environments for frontier model training.
Browse task families across five domains. Request samples and current availability.
11 collections
Long-horizon coding
Sequential engineering tasks on persistent codebases, with step-level tests for implementation quality, error recovery, and regressions.
Harbor task packages · Repository snapshots · Step-level verifiers
Long-horizon coding
Software installation, configuration, and debugging through the terminal, with executable outcome checks.
Research references
Containers · Instructions · Execution checks
Long-horizon coding
Feature development, bug fixes, and refactoring in complete repositories, evaluated with executable tests.
Research references
Repositories · Patches · Tests
Long-horizon coding
Vulnerability discovery, exploit reproduction, and patch verification in isolated software environments.
Codebases · Sandboxed targets · Regression tests
AI4Science
Expert-designed tasks for reproducing experiments, analyzing results, and validating scientific findings.
Research tasks · Reference artifacts · Expert rubrics
RSI
Controlled experiments for agents that improve models, training procedures, and their own research workflows.
Research references
Experiment environments · Trajectories · Reward signals
RSI
Research tasks in algorithm design, implementation, and empirical validation.
Research references
Challenges · Solutions · Evaluation criteria
Knowledge work
Expert-designed document and spreadsheet workflows, evaluated on accuracy, evidence, and deliverable quality.
Research references
Workspaces · Reference deliverables · Expert rubrics
Knowledge work
Multi-step investigations requiring research, tool use, and evidence-based reasoning.
Research references
Tasks · Evidence · Expert review
Agentic tool use
Agent trajectories with expert rubrics for reward modeling, critic training, and evaluation.
Research references
Session trajectories · Rubrics · Reward signals
Agentic tool use
Browser and desktop tasks requiring visual grounding, interface navigation, and cross-application execution, verified against final state.
Application environments · Screenshot–action trajectories · Outcome verifiers
Custom work
Specify a target capability or failure mode. We build the tasks, environments, and verification.
Talk to our teamBefore your training run
Work with our team to assess task difficulty, verification, and fit with your training stack.
Walk through a sampleOpen research
How well can coding agents keep building on their own code? Explore our benchmark for evolving software.
