Pretraining and fine-tuning

Pretraining teaches a model general patterns from huge corpora (often self-supervised next-token prediction).

Fine-tuning continues training on narrower data so the model behaves better for a product: chat style, domain language, tool formats, safer refusals.

You rarely train a giant LM from scratch. You adapt one. Understanding the stages explains why models know “a bit of everything,” then suddenly speak like a customer-support agent.

1. Pretrain on broad text
2. Optionally instruction-tune on examples of following directions
3. Optionally preference-tune with human or AI feedback (Course 19)
4. Ship with system prompts + tools + RAG

A general model fine-tuned on your company’s help articles and tone guide.

Write three instruction examples for a “polite librarian” assistant. That is the seed of an instruction set.

⚠️ Myth: Fine-tuning always beats good prompting + RAG.

✓ Reality: It depends on cost, update frequency, and risk; many teams mix approaches.

⚠️ Myth: Fine-tuning inserts a perfect knowledge database.

✓ Reality: It shifts behavior and weights; facts can still be wrong or stale.

• Hugging Face LLM course: https://huggingface.co/learn/llm-course/ ↗
• ANN live: https://www.ainerdnetwork.com/learn/pretraining-and-finetuning ↗
• Vaswani et al. 2017 (architecture foundation): https://arxiv.org/abs/1706.03762 ↗