Fine-Tuning Basics

Introduction

Fine-tuning updates an existing LLM's weights on your domain-specific data so it learns your tone, format and knowledge. It's the right tool when prompt engineering hits a ceiling.

Beginner analogy: The base model graduated college; fine-tuning is its on-the-job training at your company.

Understanding the topic

Core concepts to understand:

• Provide ~500–10,000 high-quality input/output pairs.
• Use LoRA / QLoRA for cheap fine-tuning of open models.
• OpenAI & Anthropic offer hosted fine-tuning APIs.
• Best for: tone, format, narrow tasks. Bad for: adding new factual knowledge.

Syntax reference

Visual workflow / architecture:

Base model (Llama 3.1 8B)
       │ + LoRA adapter
       ▼
Fine-tuned model (your tone, your format)
       │
       ▼
Inference: same speed, cheaper than GPT-4

Real-world use

Replit Ghostwriter, Harvey AI (legal), and Bloomberg GPT all rely on fine-tuned models for domain accuracy and latency.

Best practices

• Try prompt engineering first — fine-tune only when it plateaus.
• Use LoRA — 100× cheaper than full fine-tuning.
• Always evaluate against the base model.

Common mistakes

• Fine-tuning to teach facts — use RAG instead.
• Overfitting on a tiny dataset.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. When should you fine-tune vs prompt-engineer?
• Q2. What is LoRA?
• Q3. Why is fine-tuning bad at adding new knowledge?