Fine-Tuning

Fine-tuning takes a pretrained model and continues training on your dataset to shift behavior. Supervised fine-tuning learns from example prompts paired with ideal responses, and suits output format, tone and narrow classification tasks. Parameter efficient methods such as LoRA (Hu et al., 2021) freeze the base weights and train small added matrices. For GPT-3 175B the paper says this can cut trainable parameters by 10,000 times compared with full fine-tuning with Adam. Fine-tuning is a poor tool for adding facts that change (RAG is better), for raising general reasoning ability, or for patching a single bug.

Availability is the first question to ask in 2026. OpenAI is winding down its self serve fine-tuning platform: since May 7, 2026 organizations that had not fine tuned before cannot create jobs, and from January 6, 2027 active customers can no longer create new ones. Existing tuned models keep serving until their base models are deprecated. Google documents supervised tuning for Gemini. If you need a tuned model, check what your provider offers today, or tune an open weight model with LoRA and host it yourself.

import json

example = {
    "messages": [
        {"role": "system", "content": "You write short, polite refund replies."},
        {"role": "user", "content": "Where is my money?"},
        {"role": "assistant", "content": "Your refund was issued today. Banks usually post it within a few business days."},
    ]
}
print(json.dumps(example))  # one line of a JSONL training file

One training example in the chat format OpenAI documents for supervised fine-tuning. Real jobs need many lines, written or reviewed by people who know the task.

When you need consistent format or style, want to bake in domain conventions, or want to move a task to a smaller, cheaper model after a larger one has proven it. Write evals first, try prompting and RAG next, and tune only if the evals show a gap prompts cannot close. Do not tune to add facts that change, and do not start without a held out eval set.

• OpenAI: Supervised fine-tuning guide
• OpenAI: Deprecations (self serve fine-tuning)
• Google: Introduction to tuning Gemini
• Hu et al. 2021: LoRA paper

Related Terms