pierretokns/seeclick-web-commercial-mlx

SeeClick Web Commercial Dataset (MLX-VLM Format)

Commercial-use friendly GUI grounding dataset from SeeClick Web data. Apache 2.0 licensed - safe for commercial applications.

Dataset Description

This dataset contains ~20k examples for training Vision-Language Models to predict click coordinates given a screenshot and instruction. Derived from SeeClick Web crawled data (Apache 2.0).

Key Features

• License: Apache 2.0 (commercial use allowed)
• Format: MLX-VLM training format
• Examples: 19,919 total (15,935 train, 1,991 val, 1,993 test)
• Platform: Web screenshots
• Coordinate System: Absolute pixels based on actual image dimensions
• Image Resolution: Primarily 1920x1080 (coordinates scaled to actual image size)

Format

Each example follows the MLX-VLM chat format:

{
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "image"},
        {"type": "text", "text": "Click the Submit button"}
      ]
    },
    {
      "role": "assistant",
      "content": [
        {"type": "text", "text": "{\"action\": \"left_click\", \"coordinate\": [1332, 1035]}"}
      ]
    }
  ],
  "images": ["images/000123.png"],
  "image_width": 1920,
  "image_height": 1080
}

Coordinate System

• Format: Absolute pixels [x, y]
• Origin: Top-left corner [0, 0]
• Resolution: Coordinates match actual image dimensions (stored in image_width, image_height fields)
• Action: Claude Computer Use format

Important: Coordinates are scaled to the actual screenshot dimensions. Most images are 1920x1080, so a coordinate like [1332, 1035] is in the bottom-right area of the screen. Each example includes image_width and image_height fields for reference.

Usage

With MLX-VLM

# Train SmolVLM with LoRA
python -m mlx_vlm.lora \
    --model mlx-community/SmolVLM-256M-Instruct-bf16 \
    --data data/gui-grounding/seeclick-web-commercial \
    --train \
    --batch-size 2 \
    --iters 1000 \
    --learning-rate 1e-4

With Datasets Library

from datasets import load_dataset

ds = load_dataset("pierretokns/seeclick-web-commercial-mlx")

for example in ds["train"]:
    messages = example["messages"]
    image_path = example["images"][0]

Source Data

This dataset is derived from SeeClick Web, which contains:

• Screenshots crawled from public websites
• Element annotations with bounding boxes
• Apache 2.0 license (commercial use allowed)

License

Apache 2.0 - This dataset is safe for commercial use.

The original SeeClick Web data was crawled from public websites and released under Apache 2.0. No research-only components are included.

Citation

@article{cheng2024seeclick,
  title={SeeClick: Harnessing GUI Grounding for Advanced Visual GUI Agents},
  author={Cheng, Kanzhi and Sun, Qiushi and Chu, Youqi and others},
  journal={arXiv preprint arXiv:2401.10935},
  year={2024}
}