OpenAI makes it easier for developers to fine-tune and build AI models
While OpenAI is best known for its easy-to-use generative AI tools geared toward the general public, the company also has a comprehensive offering for developers and enterprises that need AI models for their specific use cases. On Thursday, OpenAI announced an expansion to its Custom Models Program and six new features for its fine-tuning API.
Fine-tuning is the lengthy process of customizing an AI model to suit specific applications and needs. Fine-tuning not only helps an AI model better accomplish the intended goal, but it can also reduce costs and lower latency.
Also: Meta's 'pruning' of Llama 2 model shows path to slimmer AI
Sometimes, the amount of customization needed exceeds what fine-tuning can provide, such as when organizations have extremely large proprietary datasets. To address this issue, OpenAI announced the Custom Model program in November 2023, letting "select" organizations get assistance from a team of OpenAI researchers to train custom GPT-4 for their specific domain.
After meeting with dozens of customers to collect feedback on how to optimize the program, OpenAI is now adding two new features: assisted fine-tuning and custom-trained models.
OpenAI says the assisted fine-tuning feature "leverages techniques beyond the fine-tuning API, such as additional hyperparameters and various parameter efficient fine-tuning (PEFT) methods at a larger scale." Assisted fine-tuning can help organizations trying to maximize their model's performance through methods such as setting up efficient training data pipelines and evaluation systems.
Custom-trained models meanwhile allow organizations to "train a purpose-built model from scratch," with the benefit of creating a model that has a deeper understanding of their business or domain.
Also: Fine-tuning OpenAI's GPT-3.5 Turbo can make it as capable as GPT-4 (if not more)
For developers who don't need as much assistance with fine-tuning, OpenAI has added features to its self-serve fine-tuning API, which lets developers fine-tune GPT-3.5 for their own needs. The new features include epoch-based checkpoint creation, comparative UI in the playground, support for integrations with third-party platforms, comprehensive validation metrics, hyperparameter configuration, and fine-tuning dashboard improvements.
OpenAI says thousands of organizations have trained hundreds of thousands of models using the API since its launch in August 2023.