Llama 2’s training corpus includes a mix of data from publicly available sources, which Meta says does not include data from Meta’s products or services. There were two trillion tokens of training ...