How an LLM becomes more coherent as we train it