Post

Introducing FACTS Grounding. A new benchmark we’re launching with Google DeepMind to evaluate LLM’s factual accuracy on over 1700 tasks. 🧠📐

We’ve partnered with Google DeepMind to publish a leaderboard of models on this new factuality benchmark. Check it out at: www.kaggle.com/facts-leader...

The leaderboard results are calculated using 1719 factual grounding tasks. Each example includes a document, a system instruction requiring the LLM to exclusively reference the provided document, and an accompanying user request.

We’ve published 860 tasks as a public data set for anyone to use: www.kaggle.com/datasets/dee...

For full details, check out the technical report & blog post. 👉Technical Report: goo.gle/FACTS_paper 👉 Blog post: deepmind.google/discover/blo...

We’ll be expanding and maintaining FACTS over time with new models and tasks. We hope FACTS Grounding helps promote the development of knowledgeable and trustworthy AI systems.