computer-use-large

license: cc-by-4.0
task_categories:
  - video-classification
  - robotics
language:
  - en
tags:
  - screen-recording
  - computer-use
  - software-tutorials
  - gui
  - desktop
size_categories:
  - 10K<n<100K
configs:
  - config_name: autocad
    data_files:
      - split: train
        path:
          - data/autocad/*
          - data/autocad_2/*
  - config_name: blender
    data_files:
      - split: train
        path:
          - data/blender/*
          - data/blender_2/*
  - config_name: excel
    data_files:
      - split: train
        path: data/excel/*
  - config_name: photoshop
    data_files:
      - split: train
        path:
          - data/photoshop/*
          - data/photoshop_2/*
  - config_name: salesforce
    data_files:
      - split: train
        path: data/salesforce/*
  - config_name: vscode
    data_files:
      - split: train
        path: data/vscode/*

Computer Use Large

A large-scale dataset of 48,478 screen recording videos (~12,300 hours) of professional software being used, sourced from the internet. All videos have been trimmed to remove non-screen-recording content (intros, outros, talking heads, transitions) and audio has been stripped.

Dataset Summary

Category	Videos	Hours
AutoCAD	10,059	2,149
Blender	11,493	3,624
Excel	8,111	2,002
Photoshop	10,704	2,060
Salesforce	7,807	2,336
VS Code	304	127
Total	48,478	~12,300

Data Fields

Each folder contains a metadata.jsonl file with the following fields per video:

Field	Type	Description
`file_name`	string	Filename of the video (e.g. `abc123.mp4`)
`category`	string	Software category
`trimmed_duration`	float	Duration of the video in seconds
`num_segments`	int	Number of contiguous screen recording segments

Data Organization

Videos are stored under data/{category}/ with a metadata.jsonl per folder. Due to HuggingFace's 10,000 file per directory limit, some categories are split across two folders (e.g. blender/ and blender_2/).

data/
  autocad/        (9,999 videos + metadata.jsonl)
  autocad_2/      (60 videos + metadata.jsonl)
  blender/        (9,999 videos + metadata.jsonl)
  blender_2/      (1,494 videos + metadata.jsonl)
  excel/          (8,111 videos + metadata.jsonl)
  photoshop/      (9,999 videos + metadata.jsonl)
  photoshop_2/    (705 videos + metadata.jsonl)
  salesforce/     (7,807 videos + metadata.jsonl)
  vscode/         (304 videos + metadata.jsonl)

Usage

from datasets import load_dataset

# Load a specific category
ds = load_dataset("markov-ai/computer-use-large", "blender")

# Load all categories
ds = load_dataset("markov-ai/computer-use-large")

Intended Use

This dataset is designed for training and evaluating computer use agents — models that interact with desktop software through GUI actions (clicking, typing, scrolling). The screen recordings provide demonstrations of real software workflows across diverse applications.

License

CC-BY-4.0