tecky-tech/Tecky-UI-Elements-VLM

Tecky UI Automation Dataset

🧠 Tecky: Your True Virtual Employee

Tecky is an AI-powered virtual teammate that lives on your machine. It learns how you interact with apps and automates workflows across both public tools and internal software.

Tecky acts as a context-aware agent, observing how users complete digital tasks and suggesting or executing them through a human-like interface.

---

📊 Dataset Overview

This dataset is the foundation for training Vision-Language Models (VLMs) to understand graphical user interfaces (GUIs) in desktop environments. It contains thousands of labeled UI elements from real-world app screenshots.

This dataset is focused solely on UI element detection and understanding.
Workflow modeling and full action-sequence understanding will be released in a future dataset.

With this data, models can:

• Parse GUI screenshots
• Identify actionable elements like buttons, inputs, sliders
• Predict appropriate click targets based on instructions
• Learn general UI semantics across diverse applications

---

🎯 Goal

The main objective is to train VLMs that:

• Understand visual structure in desktop environments
• Predict user actions (like click targets) based on screenshots and instructions
• Learn generic UI semantics across apps and layouts
• Enable automation agents like Tecky to reason and act visually

The dataset is a unified and normalized collection from open sources, focused on real desktop UI.

---

📁 Dataset Structure

Each entry in the JSONL files includes:

• system prompt (task context for Tecky)
• user message with:
  • image: path to a screenshot
  • text: instruction (e.g. “click the settings icon”)
• assistant response:
  • JSON object with:
    • actions: array of predicted clicks
    • status_update: short explanation of the action

Image paths are relative to the repo and organized in folders like images_split/00000-09999/.

We provide:

• train.jsonl: main training set
• valid.jsonl: 5% validation sample from across all sources

---

⚖️ Licenses & Attribution

This dataset merges multiple public sources. Each is preserved under its original license and properly credited.

Dataset	License	Attribution
UI element Detect Computer Vision Project by Uled	CC BY 4.0	@misc{ ui-element-detect_dataset, title = { UI element Detect Dataset }, type = { Open Source Dataset }, author = { UIed }, howpublished = { \url{ https://universe.roboflow.com/uied/ui-element-detect } }, url = { https://universe.roboflow.com/uied/ui-element-detect }, journal = { Roboflow Universe }, publisher = { Roboflow }, year = { 2023 }, month = { oct }, note = { visited on 2025-07-11 }, }
UI-Elements-Detection-Dataset by YashJain	Apache 2.0	YashJain/UI-Elements-Detection-Dataset
ui_elemenz_dataset by Maleke Chaker	MIT	Original repo
ScreenSpot by RootsAutomation	Apache 2.0	rootsautomation/ScreenSpot
Annotated UI Element Dataset for Desktop Environments	CC BY 4.0	Martínez-Rojas, A., Rodríguez-Ruíz, A., González Enríquez, J., & Jiménez-Ramírez, A. (2024). Annotated UI Element Dataset for Desktop Environments [Data set]. Zenodo. https://doi.org/10.5281/zenodo.10822752

All usage of this dataset must respect these licenses. This release falls under CC BY 4.0 to ensure attribution to all original sources.

---

🧠 Intended Use

This dataset is intended for:

• Training VLMs and UI agents (like Tecky)
• Research on UI interaction understanding
• Fine-tuning language models to act visually
• Prototyping autonomous UX testing and interface parsing

Not intended for facial recognition, surveillance, or non-UI classification tasks.

---

🙌 Contributing

Contributions welcome!

• Fork the dataset repo
• Add your JSONL/image data
• Submit a pull request
• Or open an issue with feedback or questions

---

🧾 Citation

If you use this dataset, please cite the original sources listed above and acknowledge the Tecky Project.