A Theoretical Framework for LLM Fine-tuning Using Early Stopping for Non-random Initialization