## TL;DR
Instruction tuning transforms raw language models into helpful assistants by training them to follow natural language instructions. It's the critical step between pretraining and deployment for modern LLMs.

## Core Explanation
The standard post-training pipeline: pretraining (next-token prediction on web data) → SFT (supervised fine-tuning on instruction-response pairs) → alignment (DPO/RLHF for helpfulness and safety). Instruction data sources: human-written (OpenAI), template-generated (T0, FLAN), model-generated (Self-Instruct, Evol-Instruct).

## Detailed Analysis
Instruction format: (system prompt + user instruction → assistant response). Data quality matters more than quantity — 1,000 high-quality diverse instructions often outperform 100,000 noisy ones. Evol-Instruct (WizardLM) iteratively rewrites simple instructions into increasingly complex versions.

## Further Reading
- HuggingFace: Instruction Datasets
- lmsys: Chatbot Arena
- Stanford Alpaca