AI & ML · Module

fineTuning

Fine-tuning adapts an OpenAI base model to task-specific behavior using representative input-output examples. Prepare JSONL data, upload it, create a job, and evaluate the resulting model.

How it works

Fine-tuning adapts an OpenAI base model to task-specific behavior by training it on representative input-output examples. The workflow is to prepare JSONL data, upload it, create a job, and evaluate the resulting model; reinforcement fine-tuning additionally uses a grader.

Syntax

client.fine_tuning.jobs.create(model="MODEL_ID", training_file="FILE_ID", validation_file="FILE_ID")

Parameters

Name	Type	Description
`model`	`string`	Identifier of the supported base model to fine-tune.
`training_file`	`string`	ID of the uploaded JSONL file containing training examples; the file must use the fine-tune purpose.
`validation_file`	`string`	Optional ID of the uploaded JSONL validation file used for periodic validation metrics.
`hyperparameters`	`object`	Optional hyperparameters controlling the fine-tuning job.

Returns

A fine-tuning job object containing details about the enqueued job, including status and the fine-tuned model name once available.

Key points

It specializes a base model using examples that reflect the application's expected inputs and outputs.
Begin by assembling representative training examples before starting a job.
The training dataset is uploaded to the platform in JSONL format.
Creating the job starts the model-training process after the data is available.
Reinforcement fine-tuning requires a grader that evaluates model behavior.
Evaluation is a distinct final stage used to judge whether the tuned model meets the task goal.

Common mistakes

The platform is being phased out: new users cannot access it, while existing users retain limited job-creation access and fine-tuned models remain usable until their base models are deprecated.

Example: Create a fine-tuning job

from openai import OpenAI

client = OpenAI()

job = client.fine_tuning.jobs.create(
    training_file="file-ABC123",
    model="gpt-4.1-mini",
)
print(job.id)

Creates a fine-tuning job from an uploaded JSONL training file and prints the job identifier.

When to use it

Task specialization

A task needs consistent learned behavior that prompting alone does not reliably produce.

Efficiency optimization

A smaller model should achieve the required accuracy with fewer prompt tokens.

Reinforcement fine-tuning

A behavior must be optimized against a measurable reward rather than fixed target outputs.

Iterative model improvement

A model needs evaluation-driven iteration on representative production-like examples.

Interview questions

QWhat is fine-tuning and what is its basic workflow?

Fine-tuning trains an OpenAI base model on examples representing the inputs and outputs expected in an application. The normal workflow is to collect examples, upload JSONL, create a job, and evaluate the result.

Follow-up: When should you fine-tune instead of relying only on prompting? Fine-tuning is appropriate when a specific task benefits from learned behavior; evaluation should establish whether it improves representative real-world performance.

QWhat data does a fine-tuning job require?

The training dataset is uploaded as a JSONL file, and the file is supplied when creating the fine-tuning job. A validation file is optional and produces periodic validation metrics during training.

Follow-up: Why keep validation data separate? Overlapping training and validation examples make the validation measurements less representative of generalization.

QHow does reinforcement fine-tuning differ from ordinary fine-tuning?

Reinforcement fine-tuning adds a grader that scores the model's behavior. That score guides the training process, unlike ordinary supervised examples that directly specify desired outputs.

Follow-up: What extra design problem does reinforcement fine-tuning introduce? The grader must reliably measure the target behavior, or optimization can reward undesirable outputs.

model optimization workflow
JSONL
fine-tuning job
validation metrics
reinforcement fine-tuning
grader
training file
fine-tuned model

Source: Model optimization | OpenAI API. Checked 2026-10-15.