Datasets:

Modalities:
Tabular
Text
Formats:
parquet
ArXiv:
Libraries:
Datasets
Dask
License:
rb-large / README.md
tcz's picture
Update README.md
f3ac1ce verified
|
Raw
History Blame Contribute Delete
3.63 kB
---
dataset_info:
features:
- name: svg_mobile
dtype: string
- name: svg_tablet
dtype: string
- name: svg_desktop
dtype: string
- name: html
dtype: string
- name: url
dtype: string
- name: msps_mobile
dtype: float64
- name: msps_tablet
dtype: float64
- name: msps_desktop
dtype: float64
- name: is_augmented
dtype: bool
- name: has_media_queries
dtype: bool
- name: page_sizes_desktop_width
dtype: int64
- name: page_sizes_desktop_height
dtype: int64
- name: page_sizes_tablet_width
dtype: int64
- name: page_sizes_tablet_height
dtype: int64
- name: page_sizes_mobile_width
dtype: int64
- name: page_sizes_mobile_height
dtype: int64
- name: svg_lengths_desktop
dtype: int64
- name: svg_lengths_tablet
dtype: int64
- name: svg_lengths_mobile
dtype: int64
- name: page_data_size
dtype: int64
splits:
- name: train
num_bytes: 196175064606
num_examples: 353908
download_size: 31683809119
dataset_size: 196175064606
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: apache-2.0
---
# Reverse Browser Large Dataset
This dataset contains paired HTML/CSS markup and vector-based SVG representations of over **350,000 real-world web pages**, rendered at three common screen sizes (mobile, tablet, desktop). It was collected from the public internet using a custom browser rendering pipeline. The dataset is designed for training and evaluating models that translate visual web designs to semantic, responsive markup.
## Contents
Each entry in the dataset includes:
- **HTML**: A cleaned and compressed version of the page’s markup, stripped of `<script>`, `<iframe>`, `<video>`, and other non-visual elements, only retaining `<body>` outerHTML. CSS has been inlined. Images are not saved.
- `html`: The preprocessed HTML/CSS source
- **SVGs**: Three SVG representations of the page for the following viewport sizes:
- `svg_mobile` (393×852)
- `svg_tablet` (834×1210)
- `svg_desktop` (1440×900)
- **Metadata**:
- `url`: Original webpage URL
- `msps_mobile`, `msps_tablet`, `msps_desktop`: Multi-Scale Pixel Similarity (MSPS) score for each resolution
- `page_sizes_*`: Full-length page dimensions used during rendering
- `svg_lengths_*`: File sizes of SVGs (in bytes)
- `page_data_size`: Total size of HTML + CSSs (in bytes)
- `is_augmented`: Whether the entry used data augmentation (e.g. removing or permitting elements, changing properties)
- `has_media_queries`: Whether the original markup contained responsive CSS
## Filtering Criteria
- Pages with blank screenshots (i.e., solid-color pages) were excluded.
- Entries where the **mobile** SVG’s MSPS score was lower than **0.99** were filtered out, ensuring near pixel-perfect visual fidelity between the SVG and the original screenshot.
## MSPS Metric
**MSPS** (Multi-Scale Pixel Similarity) is a custom perceptual similarity score that compares full-page screenshots and rendered SVGs. It progressively downsamples the images and computes MSE at each scale. The final MSPS score is `1 - avg(MSE)`, giving higher values to more similar images. It’s stricter than common perceptual metrics (e.g., LPIPS or SSIM) and suited for filtering high-fidelity SVG-markup pairs.
## Use Cases
This dataset supports research in:
- Image-to-Markup generation
- Responsive layout synthesis
- Visual reasoning from vector input
- Automation in web development and UI design
This dataset was produced with https://github.com/tcz/rb-crawler
https://arxiv.org/abs/2509.05394