Alignment
Domain: LLM Concepts
🎯 Motivation
Making sure the model behaves according to human values - helpful, honest, harmless. Alignment is about the model doing what humans actually want.
📋 Concrete Example
An unaligned model might give dangerous instructions when asked. An aligned model refuses harmful requests while still being helpful. RLHF is a key alignment technique.
🔗 Analogy
Raising a child with values - you don’t just teach them facts (pre-training), you teach them to be kind, honest, and safe (alignment). A knowledgeable person without values can be dangerous.