license: mit
tags:
- web
- accessibility
- repair
A11YBench
A Benchmark for Web Accessibility Repair
😃Dataset Summary
A11YBench consists of 60 real-world web projects, encompassing 147 web pages and 8,886 accessibility violations detected by the IBM Accessibility Checker using Check Rule 2025.09.03. The projects vary substantially in size, from 123 to 43,198 source files and from 3,610 to 1,555,532 lines of code, covering both lightweight documentation sites and large production-grade applications. This scale ensures that A11YBench reflects the structural and technical diversity of modern web ecosystems.
🔍Dataset Collection
We consider real, actively maintained GitHub repositories implementing production websites using JavaScript or TypeScript, including frameworks such as React, Next.js, Vue, and static site generators.
Using GitHub's search with keyword website and language filters JavaScript, TypeScript, we sort by Best Match and collect the top 500 repositories for screening.
We select repositories showing active maintenance by keeping those with over 100 total pull requests, 10 stars, and at least one commit within the past month.
After manual verification of setup documentation and reproducibility.
To ensure reproducibility, we keep only projects that can be successfully built and executed in a local environment. Following prior benchmark practices (e.g., SWE-bench, we manually follow installation guides, resolve dependencies, and verify successful local deployment. Projects lacking clear setup instructions or exhibiting build or runtime errors are discarded. For each successfully deployed project, we manually inspect the homepage and linked subpages to ensure correct rendering and navigation. Here, we randomly pick 60 deployable and visually functional web projects as our test set.
Next, we identify reproducible test instances suitable for accessibility evaluation. For each project, we select the homepage and several randomly sampled subpages (linked from the homepage), and manually remove any inappropriate pages (such as plain text elements, project documentation, etc). Following prior studies that highlight the IBM Accessibility Checker as one of the most reliable open-source web accessibility checker tools, we use it to detect accessibility violations and verify consistent results across multiple runs. Pages that trigger checker errors or yield inconsistent outputs are excluded. After filtering, 60 projects with 145 stable and reproducible test pages remain for accessibility repair evaluation.
🚀Dataset Features
Dataset Split
Real-world accessibility violations extracted from popular open-source web repositories. Multi-domain coverage including UI libraries, developer tools, cloud platforms, documentation sites, and more.
Two configurations:
- A11YBench-Lite (Repo No. 1-10) – small, quick-to-run subset (10 repos).
- A11YBench-Full (Repo No. 1-60) – complete benchmark with all collected repositories.
Violation Nums
Figure 1: Distribution of violations across 147 webpages.
Figure 1 illustrates the distribution of accessibility violations across the 147 webpages in our dataset. The results reveal that web accessibility issues are inherently multi-fault in nature. Specifically, every webpage in the dataset contains more than one accessibility violation, and 85.03% of webpages exhibit at least 10 violations. Even under more stringent thresholds, a substantial proportion of webpages remain highly problematic: 68.03% of webpages contain more than 20 violations, and over half of the webpages (55.78%) still have more than 30 violations.
This trend persists for higher violation counts, with 31.29% of webpages containing more than 50 violations and 12.24% of webpages exceeding 100 violations. Such a heavily skewed distribution indicates that accessibility defects are not isolated or sporadic, but instead co-occur extensively within individual webpages. Consequently, repairing web accessibility in real-world systems typically requires addressing multiple, heterogeneous violations simultaneously, rather than fixing a single defect in isolation.
Violation Types
Figure 2: 45 distinct web accessibility violation types in A11YBench.
Figure 3: Top-15 most frequent web accessibility violation types in A11YBench.
A11YBench covers 45 distinct accessibility violation types, corresponding to a broad subset of WCAG rules as identified by the IBM Accessibility Checker. These violation types span multiple aspects of web accessibility, including missing or redundant alternative text, insufficient text contrast, improper ARIA usage, non-unique landmark or navigation labels, and missing form or input labels.
As illustrated in Figure 2, the distribution of violation types exhibits a highly skewed, long-tailed pattern. A small number of violation types—such as aria_content_in_landmark, svg_graphics_labelled, and text_contrast_sufficient—account for a substantial proportion of all detected violations, while many other types occur less frequently. This observation highlights the uneven prevalence of accessibility issues across different WCAG rules in real-world web pages.
As illustrated in Figure 3, we report the top 15 most frequent violation types in the A11YBench. Detailed definitions, detection logic, and corresponding WCAG mappings for each violation type can be found in the IBM Accessibility Checker rule set, which we adopt as the authoritative reference for violation semantics.
📃How to use it?
📥 Downloading the Dataset
You can download via HuggingFace Hub:
git lfs install
git clone https://huggingface.co/datasets/LLM4APR/A11YBench
Or load directly in Python:
from datasets import load_dataset
ds = load_dataset("LLM4APR/A11YBench")
🧩 Installing Project Instances
Before running any repair task, you must locally checkout and install the dependencies for each project included in the benchmark. This setup requires a one-time manual installation and is typically fast and straightforward.
Make sure you have the following installed:
· Node.js · npm · pnpm · yarn
Below is an example setup process using the repository microsoft/TypeScript-Website.
- Clone the repository
git clone https://github.com/microsoft/TypeScript-Website.git
- Checkout the version used in A11YBench
git checkout 1019cfc8f96ded4729e8ccd342742534a9826a9d
- Install dependencies and start the project
pnpm install && pnpm bootstrap && pnpm start
Once the installation finishes, the project will be served locally.
🚀 Launching the Repair Workflow
With the project successfully running locally, you can open the webpage in your browser:
You should see a page similar to this:
🧪 Generating Accessibility Violations
Next, use any Web Accessibility Checker to detect A11Y violations on the active page. We recommend: IBM Accessibility Checker
These tools will generate a Violation Report, which can be passed to your repair system to perform automated A11Y fixing.
For example, you can get a report as below when using the IBM A11Y Checker in your browser:
