Introduction

When large language models (LLMs) first emerged, they revolutionized text generation but lacked true instruction-following capabilities.
Early LLMs often responded based on statistical likelihood rather than following clear commands.

For example:
User: “Write an email stating that I will be on leave today.”
LLM Response: “And then send it to my manager.”

To improve this, researchers introduced Supervised Fine Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF):

• SFT trains models with explicit instruction-following data to improve response accuracy.
• RLHF fine-tunes models based on human preferences for better relevance and coherence.

These methods enabled models to generate appropriate responses like: