Skip to content

Dataset · markov-ai

OSWorld computer agent trajectories

markov-ai/computer-use

160 successful Gemini agent trajectories on OSWorld tasks with steps and 1.0 scores.

What's inside

Contains 160 trajectories, one per OSWorld task. Total 1,378 steps, averaging ~8.6 steps per task; the agent is Gemini 3 Flash Preview with linear accessibility tree grounding. Only trajectories with a 1.0 score (fully successful) are included; examples cover domains including chrome. Modalities include image, text, timeseries, and video.

What you can do with it

  • Training computer-use agents on successful trajectories
  • Testing browser control policies
  • SFT for accessibility-tree-grounded agents
Tasks
Robotics, Image → text
Size
up to 1K rows
Downloads this month
774
Likes
77

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("markov-ai/computer-use")