showdown-clicks

General Agents

🤗 Dataset | GitHub

showdown is a suite of offline and online benchmarks for computer-use agents.

showdown-clicks is a collection of 5,679 left clicks of humans performing various tasks in a macOS desktop environment. It is intended to evaluate instruction-following and low-level control capabilities of computer-use agents.

As of March 2025, we are releasing a subset of the full set, showdown-clicks-dev, containing 557 clicks. All examples are annotated with the bounding box of viable click locations for the UI element.

The episodes range from tens of seconds to minutes, and screenshots are between WXGA (1280×800) and WSXGA+ (1680×1050). The recordings contain no PII and were collected in late 2024.

Column	Description
id	Unique identifier for each data entry (alphanumeric string)
image	Path to the screenshot image file showing the UI state
instruction	Natural language instruction describing the task to be performed
x1	Top-left x-coordinate of the bounding box
y1	Top-left y-coordinate of the bounding box
x2	Bottom-right x-coordinate of the bounding box
y2	Bottom-right y-coordinate of the bounding box
width	Width of the image
height	Height of the image

showdown-clicks-dev Results

Model	Accuracy	95% CI	Latency [^1]	95% CI
ace-control-medium	77.56%	+3.41%/-3.59%	533ms	+8ms/-7ms
ace-control-small	72.89%	+3.59%/-3.77%	324ms	+7ms/-7ms
Operator (OpenAI CUA, macOS)	64.27%	+3.95%/-3.95%	6385ms	+182ms/-177ms
Molmo-72B-0924	54.76%	+4.13%/-4.13%	6599ms	+113ms/-114ms
Claude 3.7 Sonnet (Thinking, Computer Use)	53.68%	+4.13%/-4.13%	9656ms	+95ms/-97ms
UI-TARS-72B-SFT	54.4%	+4.13%/-4.13%	1977ms	+15ms/-16ms
OmniParser V2 + GPT-4o	51.71%	+4.12%/-4.13%	12642ms	+361ms/-349ms
Gemini 2.0 Flash	33.39%	+3.95%/-3.95%	3069ms	+16ms/-16ms
Qwen2.5-VL-72B-Instruct	24.78%	+3.59%/-3.60%	3790ms	+57ms/-55ms
GPT-4o	5.21%	+1.97%/-1.80%	2500ms	+49ms/-48ms

Metrics

The evaluation script calculates the percentage of correct predictions (within the bounding box), with 95% confidence intervals created from bootstrapping.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Disclaimer

The images used in this evaluation dataset may contain content that some users might find offensive, inappropriate, or objectionable. These images are included solely for the purpose of evaluating model performance on realistic computer use scenarios.

We do not endorse, approve of, or claim responsibility for any content displayed in these images. The inclusion of any image in this dataset does not represent our views or opinions, and is not intended to promote any particular content, website, or viewpoint.

Researchers and users of this evaluation framework should be aware of this possibility when reviewing results and visualizations.

Citation

If you use showdown-clicks in your research, please cite it as follows:

@misc{showdown2025,
  title={The Showdown Computer Control Evaluation Suite},
  author={General Agents Team},
  year={2025},
  url={https://github.com/generalagents/showdown},
}

[^1]: Latency values vary significantly by provider, demand, computational resources, geographical location, and other factors - most of which are opaque to us for models we don't have direct access to. Ace models are served via General Agent's API; Qwen, Claude, Gemini, and OpenAI models utilize their respective first-party APIs; while Molmo, UI-TARS, and OmniParser models are served through Modal.