synthetic-computers-at-scale

Synthetic Computers Paper:** Synthetic Computers at Scale for Long-Horizon Productivity Simulation (arXiv:2604.28181) A dataset of **98 synthetic computer environments** designed for research on computer-use agents, long-horizon planning, and persona-grounded reasoning. Each row describes a single fictional user's computer — including the user's persona, professional context, monthly objectives, collaborators, project portfolio, filesystem policy, full file listing, and a graph of file relationships. Overview Overview of our methodology: We create user-specific synthetic computers from personas and use them as grounding environments for long-horizon productivity simulations, producing both professional deliverables and process signals for improving agents in productivity scenarios.* The collection mixes two operating systems: | OS | Computers | |---------|-----------| | macOS | 48 | | Windows | 50 | | **Total** | **98** | Quick Start Dataset Structure One row per synthetic computer. All non-id columns are **JSON-encoded strings** (use `json.loads`) — this preserves nested structure without forcing a brittle unified schema across heterogeneous personas. | Column | Type | Description | |-------------------------|--------|-------------| | `computer_id` | string | e.g. `mac_computer_000050`, `win_computer_000007`. Prefix indicates the host OS. | | `persona` | string | High-level persona description, current timestamp, OS. Keys: `persona`, `current_timestamp`, `os`. | | `user_profile` | string | Rich profile of the simulated user. Keys: `identity`, `biographical_summary`, `career_profile`, `historical_work_context`, `work_context`, `current_projects_overview`, `work_style`, `document_behavior`, `collaboration_context`, `related_people`. | | `collaboration_context` | string | Collaborators the user works with. Keys: `collaborators`. | | `monthly_objectives` | string | Simulated monthly plan. Keys: `simulation_period`, `deliverables`, `recurring_activities`, `weekly_focus`