OpenAI Launches Fine-Tuning for GPT-4o, Offering Customization and Performance Boosts

OpenAI Launches Fine-Tuning for GPT-4o, Offering Customization and Performance Boosts

August 20, 2024

OpenAI has announced the availability of fine-tuning for GPT-4o, a highly anticipated feature that allows developers to customize the model for their specific applications. This update comes with significant benefits, including improved performance, accuracy, and cost efficiency. OpenAI is also offering 1M training tokens per day for free until September 23, making it easier for organizations to get started.

What Changed Technically?

• Fine-Tuning Support: GPT-4o can now be fine-tuned using custom datasets. This process allows the model to adapt its responses and instructions based on specific use cases.
• Cost Structure:
  • Training Costs: $25 per million tokens for GPT-4o
  • Inference Costs: $3.75 per million input tokens and $15 per million output tokens for GPT-4o
  • Free Tokens: 1M training tokens per day for all organizations through September 23

Why It Matters to Practitioners

Fine-tuning is a powerful tool that can significantly enhance the performance of large language models (LLMs) like GPT-4o. Here are some key benefits:

• Customization: Developers can tailor the model's responses, structure, and tone to better fit their applications.
• Domain-Specific Instructions: The model can be trained on domain-specific data, making it more effective in specialized fields such as software engineering, creative writing, or scientific research.
• Efficiency: Fine-tuning with a small number of examples (even a few dozen) can produce strong results, reducing the need for large datasets.

Getting Started

To start fine-tuning GPT-4o, follow these steps:

1. Access the Dashboard: Visit the fine-tuning dashboard.
2. Create a New Project: Click create and select gpt-4o-2024-08-06 from the base model drop-down.
3. Upload Your Dataset: Prepare your custom dataset and upload it to the dashboard.
4. Train the Model: Use the provided training tokens or purchase additional tokens as needed.

For those interested in a lighter version, GPT-4o mini is also available for fine-tuning:

• GPT-4o Mini: Select gpt-4o-mini-2024-07-18 from the base model drop-down.
• Free Tokens: 2M training tokens per day for all organizations through September 23.

Achieving State-of-the-Art Performance with GPT-4o Fine-Tuning

OpenAI has been working with trusted partners to test fine-tuning on GPT-4o, and the results are promising. Here’s a success story:

Cosine Achieves State-of-the-Art Results on the SWE-bench Benchmark

Cosine’s Genie is an AI software engineering assistant that can autonomously identify and resolve bugs, build features, and refactor code in collaboration with users. By fine-tuning GPT-4o on examples of real software engineers at work, Cosine has achieved state-of-the-art results on the SWE-bench benchmark. This customization allows Genie to reason across complex technical problems more accurately and efficiently.

Next Steps

OpenAI is committed to expanding model customization options for developers. Fine-tuning is just the beginning, and we can expect more features and improvements in the future.

For detailed instructions and best practices, visit OpenAI’s fine-tuning documentation.