Automated End-to-End Fine-Tuning of OpenAI Models with Google Drive Integration

1. How it Works

This n8n workflow automates fine-tuning OpenAI models through these key steps:

• Manual Trigger:
  • Starts with the "When clicking ‘Test workflow’" event to initiate the process.
  • Downloads a .jsonl file from Google Drive
• Upload to OpenAI:
  • Uploads the .jsonl file to OpenAI via the "Upload File" node (with purpose "fine-tune").
• Create Fine-tuning Job:
  • Sends a POST request to the endpoint https://api.openai.com/v1/fine_tuning/jobs with:

    {  
      "training_file": "{{ $json.id }}",  
      "model": "gpt-4o-mini-2024-07-18"  
    }  
    

  • OpenAI automatically starts training the model based on the provided file.
• Interaction with the Trained Model:
  • An "AI Agent" uses the custom model (e.g., ft:gpt-4o-mini-2024-07-18:n3w-italia::XXXX7B) to respond to chat messages.

---

2. Set up Steps

To configure the workflow:

1. Prepare the Training File:

   • Create a .jsonl file following the specified syntax (e.g., travel assistant Q/A examples).
   • Upload it to Google Drive and update the ID in the "Google Drive" node.

2. Configure Credentials:

   • Google Drive: Connect an account via OAuth2 (googleDriveOAuth2Api).
   • OpenAI: Add your API key in the "OpenAI Chat Model" and "Upload File" nodes.

3. Customize the Model:

   • In the "OpenAI Chat Model" node, specify the name of your fine-tuned model (e.g., ft:gpt-4o-mini-...).
   • Update the HTTP request body (Create Fine-tuning Job) if needed (e.g., a different base model).

4. Start the Workflow:

   • Use the manual trigger ("Test workflow") to begin the upload and training process.
   • Test the model via the "Chat Trigger" (chat messages).

5. Integrated Documentation:

   • Follow the instructions in the Sticky Notes to:
     • Properly format the .jsonl (Step 1).
     • Monitor progress on OpenAI (Step 2, link: https://platform.openai.com/finetune/).

---

Note: Ensure the .jsonl file adheres to OpenAI’s required structure and that credentials are valid.