Computer Use Large: The Largest Open-Source Dataset for AI Agents

Markov AI releases 'computer-use-large', a massive dataset of 48,000+ screen recordings for training AI agents to use professional software.

by HowAIWorks Team
Markov AIComputer UseAI AgentsOpen Source DatasetHuggingFaceGUIsMachine LearningDataset

Introduction

The race to build reliable Computer Use Agents—AI models that can navigate digital interfaces just as humans do—has hit a significant milestone. Markov AI has just released Computer Use Large, the largest open-source dataset of professional computer work ever made available.

As AI agents move from simple chat interfaces to directly controlling desktop software, the need for high-quality, diverse demonstration data has never been greater. This dataset provides the necessary "trajectories" for models to learn how professional software is actually used in the real world.

A Massive Scale for Professional Workflows

The scale of Computer Use Large is unprecedented in the open-source community. Hosted on HuggingFace, the dataset comprises:

  • 48,478 screen recordings: High-quality captures of software interfaces.
  • ~12,300 hours: Over a year of continuous usage if watched back-to-back.
  • Trimmed & Focused: Every video has been stripped of audio and non-essential content like intros, talking heads, and transitions to keep the focus entirely on the GUI interaction.

Diverse Software Categories

One of the most valuable aspects of this release is its focus on "professional" software. Unlike general web-browsing datasets, this collection includes sophisticated desktop applications where precise control is required:

  • AutoCAD: Engineering and design workflows.
  • Blender: 3D modeling and animation.
  • Excel: Complex spreadsheet manipulations and data entry.
  • Photoshop: Professional image editing and graphic design.
  • Salesforce: Enterprise CRM workflows and industrial data management.
  • VS Code: Real-world programming and development environments.

The data is organized by category, making it easy for researchers to fine-tune models on specific software domains.

Built for Agentic Training

This dataset isn't just a collection of videos; it's a foundation for the next generation of Large Action Models (LAMs). By providing thousands of examples of how humans navigate nested menus, use keyboard shortcuts, and interact with complex UI elements, Computer Use Large enables:

  1. Behavioral Cloning: Teaching agents to mimic expert workflows.
  2. GUI Understanding: Improving the model's ability to "see" and interpret UI components.
  3. Workflow Evaluation: Providing a robust benchmark to test how well agents can follow multi-step professional tasks.

Conclusion

The release of Computer Use Large by Markov AI is a major win for the open-source AI ecosystem. By "democratizing" access to high-quality professional trajectories, Markov AI is lowering the barrier for developers and researchers to build agents that aren't just clever conversationalists, but capable digital workers. As we move closer to a world where AI can assist with complex CAD designs or manage enterprise sales pipelines, datasets like this will be the fuel that drives that transformation.

Sources

Frequently Asked Questions

It is a large-scale open-source dataset by Markov AI containing 48,478 screen recordings (~12,300 hours) of professional software usage for training AI agents.
The dataset covers several professional applications including AutoCAD, Blender, Excel, Photoshop, Salesforce, and VS Code.
All videos were trimmed to remove intros, outros, and transitions, and audio was stripped to focus purely on the screen recording content.
The dataset is released under the CC-BY-4.0 license, allowing for broad use and adaptation with proper attribution.

Continue Your AI Journey

Explore our lessons and glossary to deepen your understanding.