Unlock Fine-Tuning for GPT-4o: Enjoy 1 Million Free Tokens Daily Through September 23!

OpenAI has announced a new feature allowing third-party developers to fine-tune their large multimodal model (LMM), GPT-4o. This customization lets developers modify the model's behavior to better meet the specific needs of their applications or organizations.

Fine-tuning enables adjustments in tone, adherence to specific guidelines, and enhancements in technical task accuracy, proving effective even with limited data.

Developers interested in this capability can access OpenAI's fine-tuning dashboard, select “create,” and choose gpt-4o-2024-08-06 from the base model dropdown menu. This announcement follows the introduction of the smaller, faster variant, GPT-4o mini, which, while less powerful, can also be fine-tuned.

“Fine-tuning can significantly enhance model performance across various domains, from coding to creative writing,” state OpenAI technical staff members John Allard and Steven Heidel in a company blog post. “This is just the beginning; we will continue to expand model customization options for developers.”

Free Tokens Available Through September 23

OpenAI emphasizes that developers can achieve impressive results with just a few dozen training examples. To celebrate this new feature, OpenAI is offering up to 1 million free tokens per day for fine-tuning GPT-4o until September 23, 2024.

Tokens serve as numerical representations of concepts and are essential for the model's input and output processes. Developers must convert their data into tokens (tokenization) to fine-tune GPT-4o effectively, utilizing OpenAI’s tools for this process.

Ordinarily, fine-tuning GPT-4o costs $25 per million tokens, while running the fine-tuned model incurs expenses of $3.75 per million input tokens and $15 per million output tokens. For those using the smaller GPT-4o mini, up to 2 million free training tokens are available daily until the September deadline, ensuring broad access to fine-tuning features.

OpenAI’s initiative to offer free tokens comes in response to fierce competition from proprietary providers like Google and Anthropic, as well as open-source models such as Nous Research's Hermes 3, based on Meta’s Llama 3.1. However, developers using OpenAI's models benefit from the convenience of not needing to host the inference or training on their own servers; they can utilize OpenAI’s infrastructure or connect their servers via OpenAI’s API.

Success Stories Demonstrate Fine-Tuning Potential

The launch of GPT-4o fine-tuning follows extensive trials with selected partners, showcasing the potential of custom-tuned models across various sectors. For instance, AI firm Cosine achieved a state-of-the-art result of 43.8% on the SWE-bench benchmark with its fine-tuned autonomous AI engineer agent, Genie, the highest among publicly declared AI models to date.

Similarly, Distyl, an AI solutions provider for Fortune 500 companies, achieved the top ranking on the BIRD-SQL benchmark with fine-tuned GPT-4o, attaining an execution accuracy of 71.83%. The model excelled in SQL tasks, including query reformulation and self-correction.

Prioritizing Safety and Data Privacy in Fine-Tuning

OpenAI maintains that safety and data privacy are paramount while expanding customization options for developers. Fine-tuned models grant organizations full control over their data, ensuring that inputs and outputs are not used to train other models.

OpenAI has also implemented multiple safety measures, including automated evaluations and usage monitoring, to uphold compliance with its policies. However, research indicates that fine-tuning can sometimes lead to deviations from safety protocols and affect overall model performance. Ultimately, organizations must weigh the potential risks against the benefits of fine-tuning.

With the recent introduction of fine-tuning capabilities, OpenAI reinforces its vision that in the future, most organizations will develop models tailored to their industry or specific business needs. This new offering marks a significant step toward that goal, emphasizing OpenAI's commitment to enabling every organization to have its own customized AI model.