DataGemma Tackles AI Hallucinations with Real-World Data

Large language models (LLMs) are at the heart of today's AI breakthroughs, capable of sifting through massive text datasets to produce summaries, spark creative ideas, and even write code. Yet, despite their prowess, these models can sometimes deliver information that's just plain wrong, a problem we call "hallucination." It's a big hurdle in the world of generative AI.

We're excited to share some cutting-edge research that's tackling this issue head-on, aiming to curb hallucinations by grounding LLMs in real-world stats. And we're thrilled to introduce DataGemma, the first open models that link LLMs with a wealth of real-world data from Google's Data Commons.

Data Commons: A Treasure Trove of Trustworthy Data

Data Commons is like a giant, ever-growing library of public data, boasting over 240 billion data points on everything from health to economics. It pulls this info from reliable sources like the UN, WHO, CDC, and Census Bureaus. By merging these datasets into a single, powerful toolset and AI models, Data Commons helps policymakers, researchers, and organizations get the accurate insights they need.

Imagine a vast database where you can ask questions in plain English, like which African countries have seen the biggest jump in electricity access, or how income relates to diabetes across US counties. That's Data Commons for you.

How Data Commons Helps Fight Hallucination

As more folks turn to generative AI, we're working to make these experiences more grounded by weaving Data Commons into Gemma, our family of lightweight, top-notch open models. These DataGemma models are now available for researchers and developers to dive into.

DataGemma boosts Gemma's capabilities by tapping into Data Commons' knowledge, using two cool methods to improve the accuracy and reasoning of LLMs:

1. RIG (Retrieval-Interleaved Generation) amps up our Gemma 2 model by actively checking facts against Data Commons. When you ask DataGemma a question, it hunts down statistical data from Data Commons to give you a solid answer. While RIG isn't a new idea, the way we're using it in DataGemma is pretty special.

   Example query: ''Has the use of renewables increased in the world?'' applying DataGemma RIG methodology leverages Data Commons (DC) for authoritative data.

2. RAG (Retrieval-Augmented Generation) lets language models pull in extra info beyond what they've been trained on, making their answers richer and more accurate. With DataGemma, we use Gemini 1.5 Pro's long context window to fetch relevant data from Data Commons before the model starts crafting its response, cutting down on hallucinations.

   Example query: ''Has the use of renewables increased in the world?'' applying DataGemma RAG methodology showcases greater reasoning and inclusion of footnotes.

Promising Results and What's Next

Our early tests with RIG and RAG are looking good. We're seeing better accuracy in our models when dealing with numbers, which means fewer hallucinations for folks using these models for research, decision-making, or just to satisfy their curiosity. You can check out these results in our research paper.