Megadose The AI curve, compressed.

TIGER-AI-Lab/WebsiteBench

· GitHub · LLM repos ·
Agents must rebuild working websites from browser exploration, with hidden tests checking both function and visual fidelity.

WebsiteBench includes 70 websites and 46,916 hidden test cases. Each case passes only if its functional checks succeed and every visual checkpoint reaches a normalized SSIM of 0.99; a task’s reward is the share of cases passed. Tasks are available through Harbor Hub, Hugging Face and a PyPI download package. The code, task definitions and verifier use Apache-2.0, while reproduced third-party website content is restricted to non-commercial research.

GitHub · LLM repos' note

score 5

Categories: OSS & Tools, Research