The LLM training phase that produces the foundation model. Raw text, no human labels, next-word prediction at massive scale. Most of what an LLM knows comes from this phase.