OpenAI has announced an expansion of its Custom Model program, aimed at assisting enterprise clients in developing bespoke generative AI models tailored to specific use cases and domains.

Launched last year at OpenAI’s inaugural developer conference, DevDay, Custom Model initially provided companies with the opportunity to collaborate with dedicated OpenAI researchers to train and optimize models for specialized domains. Since then, "dozens" of customers have enrolled in the program. OpenAI's decision to expand the program stems from insights gained while working with these initial users, with the aim to further enhance performance.

The expansion introduces two key components: assisted fine-tuning and custom-trained models.

Assisted fine-tuning goes beyond standard fine-tuning techniques, incorporating additional hyperparameters and parameter-efficient fine-tuning methods at a larger scale. This enables organizations to establish data training pipelines and evaluation systems, enhancing model performance for specific tasks.

Custom-trained models utilize OpenAI’s base models and tools, such as GPT-4, to build tailored models for customers requiring deeper fine-tuning or domain-specific knowledge infusion. For example, SK Telecom partnered with OpenAI to fine-tune GPT-4 for improved performance in telecom-related conversations in Korean, while Harvey, supported by the OpenAI Startup Fund, collaborated with OpenAI to develop a custom model for case law, incorporating extensive legal text and expert feedback.

OpenAI envisions a future where organizations across industries develop personalized AI models to achieve more meaningful and specific impacts. This expansion aligns with OpenAI's soaring revenue, nearing $2 billion annually, and its strategic partnership with Microsoft in developing a $100 billion data center. Custom model training not only sustains revenue growth but also addresses the increasing demand for generative AI solutions.

Moreover, fine-tuned and custom models could alleviate strain on OpenAI’s model serving infrastructure, being smaller and more performant than general-purpose models. Alongside the Custom Model program expansion, OpenAI has introduced new fine-tuning features for developers working with GPT-3.5, including a dashboard for model comparison and integration support with third-party platforms like Weights & Biases. Details on fine-tuning for GPT-4, launched during DevDay, remain undisclosed.

Link: https://openai.com/blog/introducing-improvements-to-the-fine-tuning-api-and-expanding-our-custom-models-program