WebA11yBench: A Benchmark for Web Accessibility in LLM Generated Websites

A curated dataset of 5000 prompt-code pairs for fine-tuning large language models to generate WCAG-compliant HTML/CSS/JavaScript code. Each entry contains a natural language prompt describing a web feature and corresponding accessible implementation code. The dataset was distilled from DeepSeek-V3 outputs and emphasizes semantic HTML, ARIA attributes, keyboard navigation, screen reader compatibility, and responsive design patterns. Designed for training models to produce production-ready, accessibility-first web code.