OpenAI expands its custom model training program

OpenAI is expanding its Custom Model program to assist enterprise clients in creating tailored generative AI models for specific applications and domains. Launched last year at OpenAI’s DevDay conference, Custom Model has attracted “dozens” of customers, prompting OpenAI to enhance the program to optimize performance further.

The expanded program introduces two key components: assisted fine-tuning and custom-trained models. Assisted fine-tuning utilizes advanced techniques, including additional hyperparameters and parameter efficient fine-tuning methods, to improve model performance on specific tasks. Custom-trained models, on the other hand, are built using OpenAI’s base models and tools, allowing customers to deeply fine-tune models or incorporate domain-specific knowledge.

OpenAI highlights examples such as SK Telecom and Harvey, who have leveraged Custom Model to enhance GPT-4’s performance for telecom-related conversations in Korean and develop a custom model for legal tools, respectively.

With the belief that personalized models tailored to industry, business, or use case will become commonplace, OpenAI emphasizes the importance of custom model development. This initiative aligns with OpenAI’s growth trajectory, with reports suggesting it’s nearing $2 billion in annualized revenue. As the demand for generative AI continues to surge, fine-tuned and custom models offer a solution to alleviate strain on OpenAI’s model serving infrastructure.

In addition to the expanded Custom Model program, OpenAI introduces new model fine-tuning features for GPT-3.5 developers, including a dashboard for comparing model quality, third-party platform integrations, and tooling enhancements. However, details on fine-tuning for GPT-4 remain undisclosed following its early access launch at DevDay.