Google has introduced DataGemma as part of its Gemma series and released a research report, responding to the wider issue of hallucinations in large language models (LLMs).
The new feature connects LLMs with publicly available information and data from the Data Commons platform to enhance the accuracy of model outputs. In 2023, LLMs underwent rapid iterations, but issues with output stability persisted.
Fact-based LLM
The Cambridge Dictionary even selected "Hallucinate" as its word of the year, referencing this problem. The dictionary also added new entries for terms like "LLM," "prompt engineering," and "GenAI."
According to Google, while LLMs are advancing, they sometimes present inaccurate information with unwarranted confidence, commonly known as hallucinations. These errors are more frequent when querying real-time data, statistics, or numerical information.
Data Commons, a vast global database with over 240 trillion data points, sources information from trusted organizations such as the United Nations, the World Health Organization, and the US Centers for Disease Control and Prevention.
By leveraging Data Commons, DataGemma is expected to improve LLMs' factual accuracy and reasoning capabilities. This is achieved through techniques like Retrieval-Interleaved Generation (RIG) and Retrieval-Augmented Generation (RAG).
When prompted, DataGemma automatically retrieves statistics and relevant background information from Data Commons before generating an answer.
Google noted that while RIG is not a new method, its application in the DataGemma framework is unique. RAG, on the other hand, broadens the scope of possible responses to prompts.
The combination of these two techniques has led to a significant improvement in LLMs' accuracy, particularly when handling numerical and factual data.
DataGemma is primarily aimed at researchers and developers, and Google hopes that broader use of the Data Commons platform will encourage other LLMs to generate fact-based outputs. Enhancing the quality of LLM outputs is key to the wide adaptation of these models.
Ongoing tests are being conducted, and these advancements will be gradually incorporated into other models within the Gemma and Gemini series.
