What is Fine-tuning?

The problem fine-tuning solves

A pre-trained model like GPT or Llama is trained on general internet data. It can write in many styles, discuss many topics, and answer many types of questions. But it does not know:

• Your company’s internal terminology
• The exact output format your system expects
• The tone of voice your brand uses
• The regulatory constraints specific to your industry
• Domain-specific concepts that are underrepresented in public training data

You can address some of these with prompt engineering: add instructions to every prompt. But for complex domains, long instruction lists inflate costs and still produce inconsistent results.

Fine-tuning bakes the knowledge or behaviour into the model itself, so you do not need to explain it every time.

Prompt engineering vs fine-tuning

Rule of thumb: spend one week on prompt engineering first. If quality is still insufficient after exhausting prompt techniques, evaluate fine-tuning.

Types of fine-tuning

The fine-tuning workflow

Fine-tuning with the OpenAI API

Note (September 2026): OpenAI is winding down its self-serve fine-tuning platform. Since 7 May 2026 it is closed to organisations that have not fine-tuned before; since 2 July 2026 organisations that have not used a fine-tuned model in the past 60 days cannot create new jobs either; and from 6 January 2027 no one can create new jobs; already fine-tuned models keep working until their base model is retired. The example below shows the general shape of an API fine-tuning workflow; for new projects, LoRA on an open-weight model (next section) or a cloud provider’s managed fine-tuning is the realistic route.

pip install openai

import json
from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY")

# Training data format: JSONL with messages arrays
training_examples = [
    {
        "messages": [
            {"role": "system", "content": "You are a customer support agent for an Austrian fintech company. Reply in formal German."},
            {"role": "user", "content": "Ich habe eine Frage zu meiner Rechnung."},
            {"role": "assistant", "content": "Guten Tag! Ich helfe Ihnen gerne bei Ihrer Frage zur Rechnung. Bitte teilen Sie mir Ihre Kundennummer mit, damit ich Ihr Konto prüfen kann."}
        ]
    },
    # ... add 99+ more examples
]

# Save as JSONL
with open("training.jsonl", "w") as f:
    for ex in training_examples:
        f.write(json.dumps(ex) + "\n")

# Upload training file
with open("training.jsonl", "rb") as f:
    response = client.files.create(file=f, purpose="fine-tune")
    file_id = response.id

# Start fine-tuning job
job = client.fine_tuning.jobs.create(
    training_file=file_id,
    model="gpt-4.1-mini-2025-04-14",
)
print(f"Fine-tuning job created: {job.id}")

LoRA fine-tuning with Hugging Face

For open-weight models (Llama, Mistral, Qwen, Gemma), LoRA is the standard approach:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import get_peft_model, LoraConfig, TaskType
import torch

model_name = "meta-llama/Meta-Llama-3-8B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16)

# Configure LoRA: only train a small adapter
lora_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=16,               # rank: higher = more capacity, more memory
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"],  # which layers to adapt
    lora_dropout=0.1,
)

model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# Output: trainable params: 4,194,304 || all params: 8,034,476,032 || 0.05%
# Only 0.05% of parameters are being trained

The LoRA adapter saves as a small file. Apply it to the base model at inference time.

When not to fine-tune

The prompt engineering ceiling is not actually reached: Most companies that think they need fine-tuning have not exhausted prompt engineering. Few-shot examples, structured output formats, and chain-of-thought prompting often get 80% of the quality with 0% of the fine-tuning overhead.

Your task changes frequently: Fine-tuned models encode the task into weights. Changing the task requires retraining. For rapidly evolving use cases, prompt engineering stays flexible.

Your dataset has quality problems: Fine-tuning amplifies patterns in training data. A model trained on inconsistent or low-quality examples learns to be inconsistent and low-quality.

You are under EU AI Act obligations: If your use case falls under the EU AI Act’s high-risk categories, a fine-tuned model may trigger documentation, evaluation, and compliance obligations that a prompted model using a compliant third-party API does not.

What’s next

• What is Machine Learning? : The technical foundation that makes fine-tuning work
• What is a Large Language Model? : Understanding what you are actually fine-tuning
• Custom ML vs Foundation Models : When to fine-tune vs build a custom model vs prompt a foundation model
• MLOps: Getting Started : How to manage the lifecycle of fine-tuned models in production

Further reading

• OpenAI Supervised Fine-tuning Guide : Data format specs, supported base models and the wind-down notice
• OpenAI Deprecations: self-serve fine-tuning : Timeline for the fine-tuning wind-down (checked 25 September 2026)
• Hugging Face PEFT documentation : LoRA, prefix tuning, and other parameter-efficient methods
• Amazon Bedrock fine-tuning : Managed fine-tuning for Amazon Nova, Titan, Llama 3.1/3.2 and Claude 3 Haiku models