Training a modern LLM is a three-stage pipeline — each stage shapes a different aspect of the model's final behavior, and skipping a stage produces a genuinely different (and generally less useful) model.
The Three Stages
| Stage | Trains On | Produces |
|---|---|---|
| Pretraining | Massive raw text (web pages, books, code) | A "base model" — fluent, knowledgeable, but not naturally good at following instructions |
| Instruction Tuning | (instruction, ideal response) pairs | A model that follows instructions and responds in a helpful format |
| Alignment | Human preference data (which response is better) | A model tuned to be more helpful, honest, and to refuse harmful requests |
Why a Base Model Alone Isn't Enough
Prompt to a BASE model: "Explain photosynthesis"
Likely base model behavior: continues the text as if it were
scraped from the internet — might generate a list of similar
questions, or a Wikipedia-style continuation, rather than a
direct, well-formatted answer aimed at the person asking.A base model is trained purely to predict plausible next tokens — it has no built-in notion of "this is a question I should directly answer helpfully." Instruction tuning is specifically what teaches that behavior.
Cost Reality
Pretraining is by far the most expensive stage — large compute clusters running for weeks, processing trillions of tokens. Instruction tuning and alignment use far smaller, curated datasets and require a fraction of the compute, which is why fine-tuning an existing base/instruction-tuned model is realistic for a company, while pretraining a model from scratch generally is not, outside a handful of well-resourced labs.
Practical Use Case
Understanding this pipeline clarifies what fine-tuning (see Fine-Tuning) actually does: it's almost always an additional, smaller-scale training step applied on top of an already pretrained-and-aligned model — not training a model from zero.
Common Mistakes
- Assuming "training an LLM" always means pretraining from scratch — in practice, most teams that "train" a model are fine-tuning an existing one
- Skipping alignment considerations when fine-tuning a base model directly — can reintroduce unhelpful or unsafe behaviors the original alignment stage addressed
Interview Relevance
"Walk me through how an LLM like a chat assistant is trained, end to end" — the three-stage pipeline above, explained in order with what each stage fixes, is exactly the expected answer.
Practice Question
Explain why a base model (pretrained only, no instruction tuning) would likely respond poorly to "Write a professional email declining a meeting."
