Web4StyleLab/Web4Style

Web4Style

Dataset Summary

Web4Style is a curated dataset for website design understanding, website generation, style analysis, and text-based retrieval.

The current public release contains 300 examples from 300 public websites across 25 categories. Each example provides author-written annotations summarizing a site's design intent, information architecture, interaction patterns, and visual direction.

This public release is metadata-only and does not include screenshots or other third-party website assets.

What Is Included

Each row contains:

• —id: unique sample identifier
• —category: website category label
• —website: original website URL used as reference
• —brief: high-level design brief
• —content: summary of the site's information architecture and content framing
• —function: summary of the site's key interaction patterns and functional modules
• —style: summary of the site's visual style and aesthetic direction

Together, these fields describe each website at a level that is useful for style-aware modeling without redistributing the original page imagery.

Data Files

• —data/train-00000-of-00001.parquet: primary Hugging Face upload artifact
• —data/train.jsonl: line-delimited JSON export of the same records
• —scripts/prepare_release.py: reproducible conversion script from the local source JSON

Supported Tasks

Potential uses include:

• —website category classification
• —website generation
• —website design prompt generation
• —structured description generation
• —retrieval of stylistically similar websites
• —benchmarking style-aware web understanding systems

The dataset is especially suitable for experiments where the target representation is textual rather than pixel-level reconstruction.

Data Collection and Processing

The source dataset paired website screenshots with structured textual annotations. For this public release, the screenshot field was removed and only the annotation metadata was retained.

The release script performs the following transformations:

• —drops the original image field
• —preserves the original non-image field names
• —exports the cleaned records to Parquet and JSONL

Intended Use

Recommended uses:

• —academic research
• —benchmark construction
• —retrieval and ranking experiments
• —prompt engineering and evaluation for design-oriented language models

Not recommended uses:

• —reconstructing or redistributing third-party website assets
• —identity extraction or contact scraping
• —legal assertions about ownership of third-party content

Web4Style should be understood as a compact, design-focused annotation dataset rather than a comprehensive crawl of the public web.

Licensing and Rights

This repository is intended to distribute only the dataset packaging and the author-written metadata fields included in this release.

The repository does not include screenshots, logos, photos, or other third-party website assets. Website URLs are provided for reference only. Rights to third-party website content, trademarks, brand assets, and media remain with their respective owners.

The metadata fields and repository packaging are released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC-BY-NC-4.0).

Personal and Sensitive Information

The release is not intended to contain screenshots or raw copied page assets. However, source URLs may point to websites that contain public-facing company or contact information outside this repository.

This dataset is not intended for person-level profiling, identity inference, or contact extraction.

Biases and Limitations

• —The dataset contains only 300 examples.
• —Category coverage is curated and not representative of the full web.
• —The annotation fields are subjective summaries.
• —Websites change over time, so the linked sources may no longer match the original capture context.
• —The dataset is better suited to style and structure understanding than to factual verification about the linked websites.

Because the release excludes screenshots, it is not suitable for image-based benchmark tasks unless paired with separately licensed visual data.

Release Notes

• —Split count: train=300
• —Unique websites: 300
• —Categories: 25

Usage

from datasets import load_dataset

dataset = load_dataset("parquet", data_files="data/train-00000-of-00001.parquet")
print(dataset["train"][0])

Example applications include:

• —retrieving examples with similar design characteristics
• —generating structured web design briefs from category-conditioned prompts
• —studying the relationship between site function and visual direction

Publishing

This directory is structured to be pushed directly to a Hugging Face dataset repository root.

cd hf_metadata_release
git init
git add .
git commit -m "Initial dataset release"
git remote add origin https://huggingface.co/datasets/<namespace>/<repo_name>
git push origin main

Takedown and Corrections

If you are a rights holder and believe a specific record should be removed or revised, contact the maintainers with the relevant id and website.