Markov

Overview

Markov develops computer-use datasets and environments built around complete software workflows. Its public offering describes screen recordings synchronized with mouse and keyboard events and narration. That combination records both what a person sees and how they interact with the application.

The company distinguishes general computer-use supervised data from CAD environments. Its CAD examples name applications such as AutoCAD, SolidWorks, Siemens NX, SketchUp, Revit, and CATIA, with task statements, input references, reference output files, and rubrics. Broader workflow categories include browsing, coding, design, and spreadsheets.

Markov also links to publicly released computer-use and gaming recordings. Public datasets offer a way to inspect collection style, while a commercial engagement may involve different scope and terms. Named software reflects the domain of the demonstrations; it should not be read as an endorsement or partnership with each application vendor.

Why synchronized signals matter

Our assessment is that a screen recording alone can be ambiguous supervision. In an illustrative design task, a viewer may see an object change without knowing which command, shortcut, or parameter caused it. Input events and narration can help connect the observation to the action and its purpose.

Final output files add another layer. Two trajectories may look different yet produce equally valid designs, while a visually convincing result may fail a structural requirement. A useful dataset lets the learner or evaluator distinguish process variation from genuine task failure.

Evaluating fit

Inspect a full workflow rather than a short highlight clip. Check timing alignment, action resolution, narration quality, and whether the initial files are available. For CAD tasks, assess the native output and rubric with someone familiar with the engineering or design requirements. Ask how demonstrations are reviewed and whether unsuccessful attempts or corrections are retained. A pilot should evaluate transfer to new tasks and interface states, not just replay of recorded sequences. Clarify software-version dependencies, rights to the recordings and artifacts, and the intended training uses. The key value is a coherent record of meaningful work that can be connected to a verifiable outcome.