Pretraining is the first large-scale stage, learning grammar, facts, reasoning, and code from massive text corpora via next-token prediction to create a base model. Fine-tuning continues training on a smaller, domain-specific dataset to adapt style or task behavior. Instruction tuning trains the model on (instruction, response) pairs so it reliably follows tasks rather than just continuing text. RLHF then uses human preference rankings to train a reward model and optimize the LLM toward responses humans actually prefer.

Each stage solves a different problem in the pipeline from raw text to a usable assistant. Pretraining creates general capability — grammar, facts, reasoning patterns, code, general language structure — but a raw pretrained model is better at continuing text than following instructions. Fine-tuning adapts a model to a specific domain (medical documents, legal documents, support tickets, coding examples) by continuing training on smaller, targeted data. Instruction tuning specifically targets task-following behavior: given "Summarize this paragraph," the model learns to actually produce a summary rather than just predicting a statistically likely continuation of that sentence. RLHF goes a step further — humans compare pairs of model responses and rank which is better, a reward model learns from those rankings, and the LLM is optimized against that reward model to improve helpfulness, safety, conversational behavior, refusal behavior, and alignment with user intent beyond what supervised instruction tuning alone achieves. The short version: pretraining gives general knowledge, fine-tuning adapts it, instruction tuning teaches task-following, and RLHF aligns it with human preferences.