FACTS Grounding
General QA · 2024-12-17
FACTS Grounding, from Google DeepMind and Kaggle, tests whether a long-form answer to a user request about a supplied document of up to roughly 32k tokens stays fully grounded in that document. Each response is screened for eligibility and then judged for grounding by several frontier LLM judges whose verdicts are averaged; the score is the percentage of responses judged both eligible and fully grounded. Scores come from the Kaggle leaderboard over its public and private splits or from the developer's own run, as each observation notes; higher is better.