Fine-tuning API improvements and expansion of the custom models program

OpenAI has released new features for the fine-tuning API to help developers improve model performance and reduce cost and latency. This update gives developers more control, and also expands the custom models program, which involves working with experts to build models.

The newly introduced fine-tuning API features are as follows:

• Epoch-based Checkpoint Creation: Automatically generates checkpoints at each training epoch, preventing overfitting and reducing the need for retraining
• Comparative Playground: Provides a UI for comparing outputs from multiple models or snapshots at a glance
• Third-party Integration: Supports integration with external platforms such as Weights and Biases
• Comprehensive Validation Metrics: Enables loss and accuracy measurement across the entire validation dataset
• Hyperparameter Configuration: Allows direct hyperparameter configuration from the dashboard

OpenAI is also expanding the Custom Model program through its Assisted Fine-Tuning service. This is a process of working together with OpenAI researchers to build models optimized for specific domains using advanced techniques such as PEFT (Parameter Efficient Fine-Tuning).

As a real-world example, the recruiting platform Indeed fine-tuned GPT-3.5 Turbo to reduce prompt token counts by 80%, scaling monthly message processing volume to around 20 million. SK Telecom in Korea is also collaborating with OpenAI to build a telecom-domain expert model.