What is Fine-tuning? — AI Glossary

Definition

Fine-tuning is the process of taking a pre-trained model and training it further on a smaller, task-specific dataset to specialize its behavior.

A base LLM like GPT-4 or Claude is trained on the entire internet. Fine-tuning narrows that broad knowledge: you feed it hundreds or thousands of examples showing the exact input/output behavior you want, and the model adjusts its weights to match.

Common use cases: matching a specific writing style, learning domain terminology (medical, legal), or consistently outputting in a structured format.

How It Works

Pre-trained Model (general knowledge)
        ↓
  + Your Dataset (500-10,000 examples)
        ↓
  Fine-tuning Process (hours on GPU)
        ↓
  Specialized Model (your behavior)

Each training example is an input-output pair:

{"messages": [
  {"role": "system", "content": "You are a customer support agent for Acme Corp."},
  {"role": "user", "content": "How do I reset my password?"},
  {"role": "assistant", "content": "Hi! To reset your password: 1. Go to acme.com/reset 2. Enter your email 3. Click the reset link sent to your inbox. Takes about 2 minutes!"}
]}

Why It Matters

Consistency — The model reliably follows your format, tone, and rules without long prompts
Prompt savings — Fine-tuned models need shorter prompts, reducing token costs
Domain expertise — The model learns specialized terminology and reasoning
Privacy — For self-hosted models, your data never leaves your infrastructure

When to Use Fine-tuning vs RAG vs Prompting

Approach	Cost	Speed	Best For
Prompt Engineering	Free	Instant	Most tasks, start here
RAG	Low	Hours	Factual Q&A, current data
Fine-tuning	High	Days	Style, format, behavior
LoRA	Medium	Hours	Resource-efficient fine-tuning

Example

# Fine-tuning with OpenAI API
from openai import OpenAI

client = OpenAI()

# 1. Upload training data
file = client.files.create(
    file=open("training_data.jsonl", "rb"),
    purpose="fine-tune"
)

# 2. Create fine-tuning job
job = client.fine_tuning.jobs.create(
    training_file=file.id,
    model="gpt-4o-mini-2024-07-18"
)

# 3. Use the fine-tuned model
response = client.chat.completions.create(
    model="ft:gpt-4o-mini:my-org:custom-model:abc123",
    messages=[{"role": "user", "content": "How do I cancel my subscription?"}]
)

Key Takeaways

Fine-tuning changes model behavior, not knowledge — use RAG for adding facts
Start with prompt engineering; only fine-tune if prompts can’t achieve the consistency you need
You need quality training data — garbage in, garbage out
LoRA lets you fine-tune large models with much less compute
OpenAI, Google, and Hugging Face all offer fine-tuning APIs

Part of the DeepRaft Glossary — AI and ML terms explained for developers.

Fine-tuning