AI & ML · Const

Reinforcement Learning

Reinforcement fine-tuning (RFT) trains a reasoning model by grading generated responses and reinforcing higher-scoring behavior. It is intended for complex, domain-specific tasks that require advanced reasoning.

How it works

Reinforcement fine-tuning trains a reasoning model by generating answers to training prompts, grading those answers with an expert-defined grader, and reinforcing higher-scoring reasoning. The workflow uses a JSONL dataset, an uploaded training file, a fine-tuning job, and subsequent evaluation.

Syntax

client.fine_tuning.jobs.create(training_file, model, method={type: "reinforcement", reinforcement: {grader}})

Parameters

Name	Type	Description
`training_file`	`string`	JSONL training dataset uploaded to OpenAI and referenced by its file identifier.
`model`	`string`	The reasoning model to fine-tune; the page lists o4-mini-2025-04-16 for RFT.
`method`	`object`	Fine-tuning method configuration containing reinforcement fine-tuning and a grader.
`grader`	`object`	Expert-defined grader used to score generated model behavior.

Returns

A fine-tuning job object whose identifier can be used to monitor the training job and evaluate the resulting model.

Key points

RFT evaluates a generated response with an expert-defined grade and reinforces behavior associated with better scores.
Higher-scoring responses provide the training signal for improving the model’s reasoning behavior.
An RFT job includes a grader that scores the model’s behavior on generated answers.
Evaluation follows training so you can determine whether the fine-tuned model improved.

Common mistakes

A reliable expert grader is required; disagreement about the ideal result can make reinforcement signals unreliable.
This method is restricted to reasoning models rather than general-purpose models.

Example: Create a reinforcement fine-tuning job

from openai import OpenAI

client = OpenAI()

job = client.fine_tuning.jobs.create(
    training_file="file-abc123",
    model="o4-mini-2025-04-16",
    method={
        "type": "reinforcement",
        "reinforcement": {
            "grader": {"type": "string_check", "name": "exact_match"}
        },
    },
)
print(job.id)

Creates a reinforcement fine-tuning job using a training file, a supported reasoning model, and a grader.

When to use it

Specialized reasoning

Complex domain-specific tasks requiring advanced reasoning.

Medical diagnosis support

Medical diagnoses based on history and diagnostic guidelines.

Legal passage selection

Determining relevant passages from legal case law.

Fine-tuning
Supervised fine-tuning
Direct preference optimization
Evaluation
Graders

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