FACTS Grounding

FACTS Grounding is a benchmark that evaluates Large Language Models' ability to generate factually accurate responses grounded in provided documents. Featuring 1,719 diverse examples across various document lengths and user requests, this benchmark uses multiple LLM judges (Gemini 1.5 Pro, GPT-4o, Claude 3.5 Sonnet) to assess factual accuracy and grounding capabilities, measuring how well models anchor their responses to source material.