Design2Code: 484 Webpage Designs for Multimodal LLM Testing

484 webpages from the C4 validation set serve as a testbed for multimodal large language models. The dataset focuses on the task of converting visual designs into code implementations. SALT-NLP created this resource, which was last updated on March 11, 2024.

Use Cases

• Benchmarking multimodal LLM performance on visual design-to-code conversion based on webpage screenshots.
• Training models to generate HTML/CSS from visual layouts based on the described task.
• Studying the capabilities and failure modes of vision-language models on structured output tasks.

Strengths

• Contains 484 distinct webpage samples for testing.
• Sourced from the established C4 validation set.
• Explicitly designed for a specific multimodal LLM evaluation task.

Limitations

• All images in the webpages are replaced by a single placeholder image, limiting visual complexity.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

SALT-NLP

Collection Method

Derived from the C4 validation set of webpages.

Freshness

Last updated 2024-03-11 21:12:23.