A custom-made GPT-4o?

OpenAI is introducing a new customization feature, that will allow organizations to use their own data to customize, or ‘fine-tune’, its newest, most powerful Large Language Model (LLM)—GPT-4o—for specific use cases.

With this new fine-tuning capability, developers will be able to train GPT-4o on additional, custom information related to the business, so it can perform specialist, business-focused tasks, with higher accuracy, like training an employee to complete a new job or task, expertly.

For example, a company could fine-tune GPT-4o to become a specialized customer service chatbot that can answer niche, complex customer queries.

Developers can feed their datasets (just text-based data) into the model, and it will take, on average, 1-2 hours to train, meaning they can deploy customizations rapidly.

During testing, OpenAI discovered that fine-tuned GPT-4o models achieved an average accuracy level of 71.83%, and excelled at query reformulation, chain-of-thought and self-correction.

Organizations that pay for ChatGPT, will have full ownership of the customized model, the data used to train the model, and all inputs and outputs that have been generated by the finely-tuned model.

Plus, to avoid misuse, OpenAI has implemented several safety protocols including continuous automated safety evaluations and monitoring to make sure developers are following usage policies.