DeepSeek has unveiled the foundational principles of its large-scale model training, employing a meticulous two-phase approach encompassing pre-training and refined optimization. During the pre-training phase, the model leverages vast amounts of publicly accessible internet data. Conversely, the optimized training phase incorporates meticulously crafted question-answer pairs and anonymized user data, ensuring user privacy. Ultimately, the model produces text in an autoregressive fashion, seamlessly integrating these diverse training methodologies.