AI & ML · Const

Fine-tuning

Fine-tuning adapts an OpenAI base model to a specific task using representative input-output examples. The process involves preparing JSONL data, creating a job, and evaluating the resulting model.

How it works

Fine-tuning adapts an OpenAI base model to examples of the inputs and outputs expected by an application. The workflow is to collect representative data, upload it as JSONL, create a training job, and evaluate the resulting model; reinforcement fine-tuning also requires a grader.

Syntax

client.fine_tuning.jobs.create(training_file="file-...", model="...")

Parameters

Name	Type	Description
`model`	`string`	The base model to adapt; it must be one of the supported models.
`training_file`	`string`	The identifier of an uploaded JSONL file containing training examples.
`validation_file`	`string`	Optional identifier of an uploaded JSONL validation file used for periodic validation metrics.

Returns

A fine-tuning job whose eventual result includes the newly created fine-tuned model name.

Key points

It specializes an OpenAI base model using examples that represent the inputs and outputs your application expects.
Begin by assembling examples that reflect the task and production inputs.
Upload the training dataset as a JSONL file.
Start training by creating a fine-tuning job through the dashboard or API.
Reinforcement fine-tuning additionally uses a grader to assess model behavior.
Measure the trained model against representative test data after training.

Common mistakes

The platform is being phased out: new users cannot access it, while existing users can create jobs only for a limited period.

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

Specialize task behavior

A task needs consistent behavior that prompting alone does not provide.

Adapt to application examples

A production workload repeats a well-defined input-output pattern.

Optimize with a grader

A reinforcement fine-tuning workflow needs scored behavior.

model optimization
JSONL
training file
validation file
reinforcement fine-tuning
grader
evaluation

Source: Model optimization. Checked 2026-10-15.