Model Distillation feature launched via API

OpenAI has launched the Model Distillation feature, providing an integrated workflow that lets developers fine-tune cost-efficient models such as GPT-4o mini using the outputs of high-performance models like o1-preview or GPT-4o.

Previously, the model distillation process was cumbersome, requiring developers to manually link multiple tools across dataset creation, fine-tuning, and performance measurement, but now the entire process can be managed within the OpenAI platform. The key features introduced this time are as follows.

• Stored Completions: Automatically stores input-output pairs generated via the API to build high-quality fine-tuning datasets. Data can be collected without added latency using the store:true flag.
• Evals (beta): An integrated tool for creating and running custom evaluations to measure model performance on specific tasks. Developers can upload existing datasets or use Stored Completions data.
• Fine-tuning: Fully integrated with the existing fine-tuning service, allowing immediate fine-tuning using datasets built with Stored Completions and performance verification with Evals.

Developers can optimize model performance through an iterative process: first establishing a baseline with Evals, then building a dataset with Stored Completions, and using it to fine-tune the model.

The feature is currently available to all developers, and to support adoption, OpenAI is providing free daily training tokens until October 31st — 2 million tokens per day for GPT-4o mini and 1 million tokens per day for GPT-4o.