The training corpora of large language models typically include Common Crawl web data, books, academic papers, open-source code repositories, and content from Q&A platforms such as Stack Overflow, with knowledge compressed into parameters through self-supervised learning (predicting the next token).