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

Data Fields

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

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