Pre-training versus fine-tuning an LLM

WHAT THIS TESTS The interviewer wants to confirm you understand that an LLM is built in stages with very different data, cost, and goals, rather than one monolithic training run.

A GOOD ANSWER COVERS Pre-training is self-supervised: the model predicts the next token over trillions of tokens of unlabeled web text, books, and code. The objective is generic language modeling, so no human labels are required, but the compute is enormous, often thousands of GPUs over weeks and a large fraction of total cost. Fine-tuning starts from those frozen-then-updated weights and adapts them on a much smaller, curated dataset, frequently labeled or instruction-formatted. The objective is narrow: follow instructions, match a domain style, or solve a specific task. Cost is orders of magnitude lower, and techniques like LoRA reduce it further by updating only a few parameters.

COMMON WRONG ANSWERS Saying pre-training uses labeled data, conflating the two objectives, or claiming fine-tuning injects large amounts of new factual knowledge. Fine-tuning mostly steers and surfaces capabilities already learned during pre-training; it rarely teaches genuinely new facts reliably.

LIKELY FOLLOW-UPS Expect questions on parameter-efficient fine-tuning such as LoRA, on instruction tuning versus continued pre-training, on catastrophic forgetting, and on when retrieval augmentation is preferable to fine-tuning for adding knowledge.

ONE CONCRETE EXAMPLE A base model trained on web-scale text learns grammar, facts, and reasoning patterns through next-token prediction. A startup then fine-tunes it on five thousand customer-support conversations so it answers in the company tone and follows the support playbook. The base run might cost millions of dollars; the fine-tune runs in hours on a single machine, yet it does not add the company's internal pricing tables, which are better supplied via retrieval at inference time.