OpenAI Fine-tuning for JSON Output

Introduction

This project, written in JavaScript (Node.js), fine-tunes OpenAI's primary models to produce JSON-formatted output. It aims to enhance both cost-efficiency and data consistency for applications that utilize the OpenAI API. Portions of the codebase are inspired by the Python open-finetuning.

Prerequisites

• Node.js 16 LTS or later
• OpenAI Node v4

Usage

Step 1: Set OpenAI API Key Environment Variable

Define your OpenAI API key by setting an environment variable named API_KEY.

export API_KEY=your-openai-api-key

Step 2: Generate Sample Training Data

• Execute npm run prepare.
• Use a prompt capable of generating the desired output.
• For optimal sample quality, it is recommended to use GPT-4.
• OpenAI requires a minimum of 50 samples.
• The sample data will be stored in src/data/training-examples.json.

Step 3: Generate System Message Formatting Data

• Execute npm run transform.
• Use a more straightforward prompt than the previous step.
• The results will be saved in src/data/training-data.jsonl.

Step 4: Proceed with Fine-tuning

• Execute npm run finetune.
• Upon completion, the fine-tuned model ID will be displayed on the console.

Step 5: Test the Fine-tuned Model

• Execute npm run model.
• Test the model using a simplified version of the original prompt.