“The resulting dataset contains 2.7 million unique public domain U.S. newswire articles, written between 1878 and 1977.”
Researchers ran hundreds of terabytes of local newspaper scans through a deep learning pipeline and transcribed around 138 million article texts. They kept only the newswire copy, because that is the part in the public domain. The pitch lists language model training first and history second. A century of journalism is now more valuable as training data than as a record.