OpenAI Enables Custom Fine-Tuning for GPT-3.5 Turbo
OpenAI now allows developers to fine-tune GPT-3.5 Turbo using custom data to improve reliability and match GPT-4 on specific tasks. The update also helps businesses reduce prompt sizes and lower API costs.
OpenAI introduces a fine-tuning feature for GPT-3.5 Turbo, allowing customers to train the model with custom data to improve its reliability and build specific behaviors. The company states that on certain narrow tasks, a fine-tuned GPT-3.5 Turbo matches or even outperforms the base capabilities of GPT-4, giving developers a powerful tool to create unique user experiences at scale.
This new capability enables businesses to make the AI model better follow instructions, maintain a consistent response format for tasks like code completion, and adopt a specific brand tone. Additionally, fine-tuning allows customers to significantly shorten their text prompts, with early testers reducing prompt sizes by up to 90%, which directly speeds up API calls and cuts operational costs.
To use the feature, developers currently prepare and upload data files to create fine-tuning jobs through the API, with all data undergoing strict moderation checks. OpenAI charges specific rates per token for training and usage, making a typical 75,000-word training job cost around $2.40, and plans to release a dedicated fine-tuning user interface dashboard in the future.