Markov

Training AI models to use computers requires large volumes of data that capture real human interactions with software, including mouse and keyboard events, narrations, and verifiable task outcomes. This kind of data is difficult to produce at scale across the many domains computer-use AI needs to handle, from CAD tools to browsers to productivity software. Without synchronized interaction signals and complete workflows tied to problem statements and gold outputs, model builders lack the structured training material needed to teach AI systems to perform real computer tasks reliably.

Markov captures how people use computers and turns those sessions into structured datasets for training computer-use AI. Its data spans CAD environments (AutoCAD, SolidWorks, Siemens NX, SketchUp, Revit, CATIA) as well as browser use, coding, design, and spreadsheet workflows, each packaged as screen recordings with synced mouse and keyboard events and narrations. For CAD workflows specifically, datasets also include problem statements, input references, gold output files, and rubrics to support supervised fine-tuning. Markov also publishes open-source datasets on Hugging Face under markov-ai, including over 10,000 hours of computer-use recordings across tools like Salesforce, Blender, and Photoshop, and over 500 hours of gaming data with synchronized inputs from games such as Valorant, Minecraft, and GTA.