OpenAI Enables Custom Fine-Tuning for GPT-3.5 Turbo Model
OpenAI now allows developers to fine-tune GPT-3.5 Turbo with custom data to improve reliability and reduce prompt sizes. The company claims these customized models can match or beat GPT-4 on specific, narrow tasks.
OpenAI introduces fine-tuning capabilities for its GPT-3.5 Turbo model, allowing developers to train the AI with custom data. This new feature helps businesses improve the model's reliability, ensure consistent formatting for tasks like code completion, and adjust the tone to match a specific brand voice. OpenAI states that a well-fine-tuned GPT-3.5 Turbo model matches or even outperforms its flagship GPT-4 model on certain narrow tasks.
The fine-tuning process significantly cuts costs and speeds up API performance by reducing prompt sizes. Early testers shrink their prompt lengths by up to 90% by baking instructions directly into the model. To maintain safety, all custom training data passes through OpenAI's moderation API and a GPT-4-powered moderation system before a fine-tuning job begins.
Developers currently set up fine-tuning through the API by preparing and uploading data files, though OpenAI plans to release a user interface dashboard in the future. The pricing for this feature breaks down to $0.008 per 1,000 tokens for training, $0.012 for input usage, and $0.016 for output usage. Alongside this update, OpenAI releases two updated GPT-3 base models, babbage-002 and davinci-002, while scheduling the retirement of the original GPT-3 base models for January 4, 2024.