Google Bard is based on the LaMDA language model, which was trained on Infiniset datasets based on Internet content.
Only 12.5% of the data used to train LaMDA AI comes from a public dataset of crawled web content, and another 12.5% comes from Wikipedia, according to the 2022 LaMDA research paper.
Google Bard is built on the LaMDA language model, which stands for Language Model for Dialogue Applications.
LaMDA was trained using the Infinite dataset. It was pre-trained on 1.56 trillion words of "public dialogue data and web text" in total.
Only 25% of the data comes from a specific source (the C4 dataset and Wikipedia). The remaining data accounts for 75% of the Infiniset dataset & consists of words scraped from the Internet.
However, it's yet to know how the data is been obtained, which websites were used, or any other information about the scraped content.
The best word to describe the 75% of data that Google used to train LaMDA is murky.
Some hints may provide a general idea of what sites are contained within 75% of web content, but we can't be sure.