Datasets:

Modalities:
Image
Text
Formats:
parquet
Languages:
English
ArXiv:
Libraries:
Datasets
Dask
License:
WebSight / README.md
HugoLaurencon's picture
Update README.md
b11f817 verified
metadata
language:
  - en
license: cc-by-4.0
size_categories:
  - 1M<n<10M
pretty_name: WebSight
dataset_info:
  - config_name: v0.2
    features:
      - name: image
        dtype: image
      - name: text
        dtype: string
      - name: llm_generated_idea
        dtype: string
    splits:
      - name: train
        num_bytes: 368943620718.125
        num_examples: 1922671
    download_size: 144861710051
    dataset_size: 368943620718.125
  - config_name: v0.1
    features:
      - name: image
        dtype: image
      - name: text
        dtype: string
    splits:
      - name: train
        num_bytes: 35386660486.65
        num_examples: 822987
    download_size: 31394170440
    dataset_size: 35386660486.65
configs:
  - config_name: v0.2
    default: true
    data_files:
      - split: train
        path: v0.2/train-*
  - config_name: v0.1
    data_files:
      - split: train
        path: data/train-*
tags:
  - code
  - synthetic

Dataset Card for WebSight

Dataset Description

WebSight is a large synthetic dataset containing HTML/CSS codes representing synthetically generated English websites, each accompanied by a corresponding screenshot.

This dataset serves as a valuable resource for tasks such as generating UI codes from a screenshot.

It comes in two versions:

  • v0.1: Websites are coded with HTML + CSS. They do not include real images.
  • v0.2: Websites are coded with HTML + Tailwind CSS. They do include real images.

Essentially, here are the improvements in version v0.2, compared to version v0.1:

  • Websites include real images (related to the context of the website!!)
  • Usage of Tailwind CSS instead of traditional CSS
  • Contains 2x more examples
  • Contains more tables
  • Better resolution for the screenshots
  • Presence of a column indicating the LLM-generated idea to create the websites
Details for WebSight-v0.1 (HTML + CSS)

Data Fields

An example of a sample appears as follows:

{
    'images': PIL.Image,
    'text': '<html>\n<style>\n{css}</style>\n{body}\n</html>',
}

where css is the CSS code, and body is the body of the HTML code. In other words, the CSS code is embedded directly within the HTML code, facilitating the straightforward training of a model.

Data Splits

There is only one split, train, that contains 822,987 images and codes.

Dataset Creation

This dataset was created using Mistral-7B-v0.1 to generate random website ideas with the prompt

Generate diverse website layout ideas for different companies, each with a unique design element.
Examples include: a car company site with a left column, a webpage footer with a centered logo.
Explore variations in colors, positions, and company fields.
Don't give any explanations or recognition that you have understood the request,
just give the list of 10 ideas, with a line break between each.

which were then passed to Deepseek-Coder-33b-Instruct with the prompt

Create a very SIMPLE and SHORT website with the following elements: {idea}
Be creative with the design, size, position of the elements, columns, etc...
Don't give any explanation, just the content of the HTML code `index.html` starting with `<!DOCTYPE html>`,
followed by the CSS code `styles.css` starting with `/* Global Styles */`.
Write real and short sentences for the paragraphs, don't use Lorem ipsum.
When you want to display an image, don't use <img> in the HTML, always display a colored rectangle instead. 

Following these steps, the HTML and CSS codes were extracted from the outputs of Deepseek-Coder and formatted into the structure '<html>\n<style>\n{css}</style>\n{body}\n</html>'.

Details for WebSight-v0.2 (HTML + Tailwind CSS)

Data Fields

An example of a sample appears as follows:

{
    'images': PIL.Image,
    'text': '<html>\n<link href="https://cdn.jsdelivr.net/npm/[email protected]/dist/tailwind.min.css" rel="stylesheet">\n{body}\n</html>',
}

where body is the body of the HTML code, containing directly the Tailwind CSS code and facilitating the straightforward training of a model.

Data Splits

There is only one split, train, that contains TO DO images and codes.

Dataset Creation

TO DO.

For now, the creation of the dataset is documented in the technical report.

Terms of Use

By using the dataset, you agree to comply with the original licenses of the source content as well as the dataset license (CC-BY-4.0). Additionally, if you use this dataset to train a Machine Learning model, you agree to disclose your use of the dataset when releasing the model or an ML application using the model.

Licensing Information

License CC-BY-4.0.

Citation Information

If you are using this dataset, please cite our technical report

@misc{laurençon2024unlocking,
      title={Unlocking the conversion of Web Screenshots into HTML Code with the WebSight Dataset}, 
      author={Hugo Laurençon and Léo Tronchon and Victor Sanh},
      year={2024},
      eprint={2403.09029},
      archivePrefix={arXiv},
      primaryClass={cs.HC}
}