Intelligence is learned in the doing. We build the worlds where models practise, and record how experts move through them: each step, each misstep, each recovery.
Environments and trajectories, across six data families.
Standard datasets or custom builds. Every family ships with documentation, a quality report and a rights memo.
3D & Blender Agent Data
3D assets, scenes and Blender workflows, paired with code, GUI traces and RL signals.
- a3D scenes & assets
- bAsset/Scene-to-Code
- cBlender GUI trajectories
- dAgent traces & tool calls
- eTask–verifier pairs
- fError-recovery traces
Spatial reasoning, world models, Code-to-3D, computer-use and tool-use agents, RL post-training.
Scenes include object hierarchy, camera, lighting and spatial-relation labels. Agent data covers Blender API, Python and MCP tool use, including failed attempts and recoveries.
Game & Interactive World Data
Gameplay, engine and code data for agents that perceive, plan and act in rich worlds.
- aGameplay video + synchronized input
- bState–action trajectories
- cGoal-directed task demos
- dGame RL data
- eEngine projects & Code-to-Game
- fBuild & debug traces
Vision-action models, imitation learning, game RL, engine and coding agents.
We score titles for training value, not popularity: open worlds, driving, traversal, crafting, crowds, physics and long-horizon missions. Capture scales from niche titles to large open-world AAA programs. PC and console today, mobile in progress.
Code-to-Video & Demonstration Data
Videos and visual deliverables built through code and tools, with the full execution trace.
- aCode-to-Video
- bVideo-driven 3D blockout replication
- cFrame-by-frame workflows
- dDemo construction
- eVisual workflow trajectories
Code generation, creative coding, computer-use and multimodal agents.
Generated through code, graphics tools and procedural pipelines, not video-generation models, so every frame is traceable to the instructions that made it.
AI for Science Data
Scientific tasks, workflows and code with tool use, outputs and validation.
- aScientific tasks
- bResearch workflows
- cScientific coding
- dOpen-science benchmark tasks
Research agents, scientific reasoning, scientific coding agents.
Sourced and licensed task by task; scope confirmed during evaluation.
Agentic Environments & RL
Executable environments with real toolchains, multi-step tasks, verifiers and reward signals.
- aTerminal coding & tool-use
- bRepo-level SWE tasks
- cPaper reproduction & deep research
- dEnterprise workflow environments (finance, insurance, manufacturing, RevOps, legal)
Coding, SWE, research and enterprise agents; capability evals; verifiable-reward RL.
Tasks are tagged by language, codebase size, difficulty (up to Frontier), context type and checkpoints. Enterprise environments include databases, large document sets, dozens of tools and simulated users, tiered from 3 to 20+ steps.
Expert STEM Reasoning & Evaluation
Expert-written, expert-reviewed reasoning and evaluation data with verifiable answers.
- aBiomedical & clinical reasoning
- bCross-disciplinary STEM
- cAdvanced mathematics
- dProfessional knowledge benchmarks
Frontier reasoning evaluation, math and formal RL, professional-agent evaluation.
Covers math, physical and life sciences, medicine, engineering, CS, economics, linguistics, history and philosophy.
Every episode is a full record, not a screenshot.
Agent and RL datasets carry the task, every action and observation, intermediate state, the verifier and its result. Failed attempts and recoveries are kept and labelled.
- task
- Natural-language goal, input project and constraints.
- observation
- Screenshots, UI state, scene graph or terminal output per step.
- action
- Tool calls, code, mouse and keyboard events with timestamps.
- state_diff
- Before and after records of project and object state.
- verifier
- Executable grading script, rubric and scoring rules.
- reward
- Outcome label plus staged checkpoint scores.
- recovery
- Error signal, diagnosis and corrective action when a step fails.
Need something specific? We'll build it.
Bespoke production around your target environment, task taxonomy, modalities, quality bar and volume, from pilot to large-scale programs.
Tell us what you need- AGameplay capture
- BEngine & pipeline data
- CBlender GUI and RL data
- DCode-to-video workflows
- ECustom verifiers and rubrics
What every delivery includes.
Dataset overview
Scope, taxonomy, use cases, coverage.
Sample pack
Authorized samples with schema and field docs.
Technical spec
Format, versioning, environment, reproducibility.
Quality report
Distribution, error and duplication rates, QC method.
Rights memo
Provenance, licenses, usage and redistribution basis.
Clear commercial terms
Scope, exclusivity, acceptance criteria.
Agent and RL datasets can also include difficulty stratification, trajectory statistics, reward and verifier design, action schemas and execution logs.
From sample pack to full delivery.
- 01
Request samples
Tell us what you're training.
- 02
Evaluate
Your team tests the sample pack.
- 03
Scope
Taxonomy, volume, terms.
- 04
Pilot
Delivery and acceptance testing.
- 05
Scale
Full delivery and expansion.
Built for teams training agents.
Frontier AI labs
Post-training, RL and capability evaluation for coding, research and computer-use agents.
Robotics & embodied AI teams
3D scenes, spatial data and interactive-world trajectories.
Large technology companies
Enterprise-agent environments and domain-expert evaluation.
See the data before you decide.
Tell us what you're training. We'll send a relevant sample pack and dataset overview.
Prefer email? team@nexlin.ai
Thanks.
We'll reply within business days with a relevant sample pack.