Data Coverage
Defines the measurable breadth of tasks, tools, environments, and workflows represented in the
General Data platform.
Coverage is not a function of volume.
It is a function of how much of the real interaction space is captured in a form that is learnable,
reproducible, and aligned with actual human execution
Scale Metrics
The dataset is built to operate at large-scale model and agent training.
● 200K+ computer-use trajectories
● ~8M state-action pairs
● 1,500+ hours of real human data
● 130+ software tools across domains
● Average ~30 steps per task (long-horizon bias)
Task Creation
Feature-level completeness alone is a weak proxy for learning. Many features are rarely used,
weakly tied to outcomes, or do not require meaningful reasoning. Optimizing for feature
coverage only, leads to shallow interaction surfaces.
Coverage is built by breaking tools into foundational blocks and features:.
● Function Mapping defines how users interact with the UI (navigation, layout, interaction
patterns)
● Feature Mapping defines what the tool enables (its core capabilities or use cases)
They are covered through tasks at different levels of task complexity.
Tasks are divided into four types:
● Elementary (1 step): single UI interactions covering basic actions.
● Atomic (2–3 steps): short sequences covering simple feature usage
● Multi-Step (4–15 steps): combines multiple features to complete a specific goal
● Workflow (15+ steps): long-horizon execution with retries, backtracking, and context
switching
● Games: : interactive tasks with dynamic state changes and variable-length execution
Reasoning traces are aligned with these tasks to capture intent-based decision-making across
steps.
Outcome
● UI interaction space is fully covered
● Tool capabilities are exercised across increasing complexity
● Tasks capture both execution and reasoning behavior
Coverage
Each trajectory corresponds to a goal-completion, where a user moves from an intent to a
verifiable outcome. Tasks are distributed on the core function of user intent:
● Creation - generating documents, designs, structured content
● Modification - editing, formatting, transforming existing artifacts
● Retrieval - search, navigation, information extraction
● Configuration - settings, permissions, system adjustments
● Transactional - booking shortlisting, and, submissions
● Communication - messaging, scheduling, collaboration
● Analytical - working with structured data, dashboards, reporting
Diversity
Task diversity emerges naturally throughout the data collection process from variation across
multiple axes:
● Execution
Multiple valid approaches exist for the same task, with operator-specific strategies,
action ordering, and variation in decision paths and intermediate steps.
● Cross-Tool trajectories
Tasks span multiple tools and interfaces, requiring transfer of context, data, and
documents across environments, often forming workflows that cannot be completed
within a single tool.
● Environment
Tasks are executed across different screen resolutions, device configurations, operating
systems, UI layouts, themes (light/dark), and dynamic UI states such as popups, async
loading, and conditional rendering.
● Context
Execution begins from non-canonical starting states, including pre-existing files,
sessions, and dashboards, where prior context and system conditions directly influence
task behavior.
● Interaction Diversity
Tasks involve varied input modalities including mouse, keyboard, and gestures, with
differences in timing, latency, execution speed, and natural inclusion of retries,
backtracking, and corrections.
Domain Coverage
Domain coverage defines where tasks occur and how interaction patterns are distributed across
real software ecosystems. Percentages are normalized against total task volume (~161K
tasks).
Domain Tasks (% Tool Tools
of Total) Count
Office Productivity 87,507 40+ Apple Calendar, Apple Keynote, Apple Mail, Apple Numbers,
(~54.2%) Apple Pages, Asana, Confluence, Dropbox, Eval TaskSet,
Evernote, Gmail, Google Calendar, Google Classroom, Google
Docs, Google Drive, Google Forms, Google Keep, Google Meet,
Google Sheets, Google Slides, iCloud, [Link], MS Excel,
MS PowerPoint, MS Word, Multi-Tool, Notion, OneDrive,
OneNote, Outlook, Outlook Calendar, Power BI, Seafile, Skype,
SurveyMonkey, Trello, VS Code Theme Studio, Web Browsing,
WeTransfer, WinRAR, WinZip, Zoho Sprints, Zoom
Design & Creativity 14,459 7 Adobe Photoshop, Affinity Designer, Blender, Canva, CapCut,
(~8.9%) DaVinci Resolve, Figma
Coding & 13,149 16 Android Studio, Bitbucket, Docker, GitHub, Google Colab,
Development (~8.1%) Jenkins, Jira, Jupyter Notebook, Looker Studio, MDN Docs,
MySQL, PyCharm, PythonAnywhere, Sublime Text, VS Code
Communication 6,511 5 Discord, Microsoft Teams, Slack, Teamwork, WhatsApp
(~4.0%)
CRM & Operations 5,308 4 HubSpot, Klaviyo, Salesforce, Zoho CRM
(~3.3%)
Shopping / E- 6,905 9 Amazon, Ikea, Instacart, Nike, Sephora, Walmart, Zara
commerce (~4.3%)
Travel & Booking 6,894 5 Airbnb, [Link], Expedia, Google Flights, Skyscanner
(~4.3%)
Browser & Web 4,378 7 Brave, Google, Chrome, Edge, Firefox, Opera, Safari
(~2.7%)
Web Design / CMS 4,626 6 Carrd, Squarespace, Tailwind UI, Webflow, Wix, Wordpress
(~2.9%)
Other Tools 5,768 7 7-Zip, Calculator, Calendly, ChatGPT, Mailchimp, Preview, VLC
(~3.6%) Media Player
Social Media 2,237 6 Facebook, Imgur, Instagram, LinkedIn, X, YouTube
(~1.4%)
OS Control 1,679 3 macOS System Controls, Windows 10, Windows System
(~1.0%) Controls
Games 1,377 16 8 Ball Pool, Bridge, [Link], Go, Go (Game), Happy Glass,
(~0.85%) Ludo, Mario, Plant vs Zombies, [Link], Smash Karts, Spider
Solitaire, Subway Surfers, Sudoku, Temple Run, Tic Tac Toe,
Water Sort, Wordle
Maps 318 (~0.2%) 2 Apple Maps, Google Maps
Cloud / Infra 112 3 AWS, Azure, GCP
(~0.07%)
Benchmark 338 4 Mind2Web, OSWorld, SheetBench, WebVoyager
(~0.21%)
Fig. Task Distribution Across Domains (%)
Dataset Signals
● ~54% of total tasks are concentrated in office productivity workflows, reflecting the
highest-density real-world interaction environments.
● Top 5 domains account for ~75%+ of total task volume, ensuring repeated exposure to
high-frequency workflows rather than sparse distribution.
● 100+ tools across 16 domains are covered, providing broad interface diversity across
software ecosystems.
● ~20%+ of tasks come from structured domains (coding, CRM, design), where execution
involves dependencies, multi-step reasoning, and planning.
● ~10%+ of tasks come from dynamic domains (shopping, travel, social), introducing non-
deterministic UI behavior, layout shifts, and content variability.
● ~4%+ of tasks are explicitly categorized under system-level and browser interactions,
capturing navigation, control flow, and environment transitions. These interaction
patterns are also implicitly present across the broader dataset through web-based and
desktop workflows.