Training data and environments for frontier models
License what we have already built, or have it built for your model.
Capabilities
Agentic coding
Long-horizon software tasks in real repositories, worked in the terminal, from bug fixes to full-stack features, each with a verified outcome.
Request data
Computer and browser use
Desktop and web tasks across real applications, from office suites and creative tools to booking sites, with every end state checked.
Request data
Professional work
Finance, legal and consulting work as professionals actually do it: multi-app workflows and deliverables graded to an expert's standard.
Request data
Recursive self-improvement
AI research tasks for agents: reproducing papers, running experiments and post-training models, scored on whether the results hold up.
Request data
Games and 3D
Playable games, interactive 3D scenes and manufacturable parts, judged on how they look, how they play and whether they work.
Request data
Mobile app development
Complete iOS, Android and React Native apps and feature work, held to the polish people expect from the apps they use every day.
Request dataHow to buy
Off the shelf or built for you
Off-the-shelf data
Validated datasets and environments, ready to license today, with samples available before you commit.
Browse the catalogCustom datasets and environments
Datasets and environments built around your model's gaps and difficulty level, delivered at the scale of a training run.
Request custom dataEvaluations
Private benchmarks grounded in real work, re-runnable on every checkpoint so progress is measured, not assumed.
Talk to us about evalsDelivered in the form your training needs
- RL environments
- Self-contained environments built on real software, APIs and tools, where agents do multi-step work and every outcome is checked.
- SFT and reasoning traces
- Expert-quality demonstrations with full reasoning, written to teach models how to plan and work through complex tasks.
- Preference data
- Carefully ranked comparisons that capture what makes one answer better than another, for RLHF and DPO.
- Methods we support
- RL · RLHF · RLSF · GRPO · SFT · DPO · (M)OPD · SDFT · OPSD · PivotRL · and more
Data catalog
Request access to browse our off-the-shelf datasets across every capability above.
Request access