OpenAI just announced the launch of a fine-tuning feature for its GPT-3.5 Turbo, providing artificial intelligence developers with the ability to boost performance on tasks with dedicated data. The company announced this update on Tuesday and it follows the customization update in July.
The new update also allows developers to customise models that are perfect for case studies and run these models at the required size to solve the problem. This way, developers and businesses can adjust their user experiences with the tool to their use cases.
With the fine-tuning feature, developers can use their data and applications to train GPT-3.5 Turbo using the company’s API. According to the company said, ” Fine-tuning lets you train the model on your company’s data and run it at scale.”
The GPT-3.5 fine-tuning process
Companies that want to take on the fine-tuning process have to prepare and upload their data through OpenAI’s API, and then create the fine-tuning work. Once it is completed, the model can be used for subsequent jobs.
Using the fine-tuned model will cost 0.012 dollars per 1000 input token and 0.016 dollars per 1000 output token. This price is more expensive than the basic cost of 0.0004 per 1000 tokens for the GPT-3.5 Turbo. Companies may need to spend additional costs for the training process, depending on how large the data is.
The company had a private beta-testing of the products and the tests suggest that a fine-tuned GPT-3.5 Turbo version can perform better than GPT-4 on narrow tasks
Safety procedures
OpenAI will transfer the data through its GPT-based moderation system and moderation API to ensure safety and adherence to safety standards. With this, OpenAI can control the kind of data that developers and companies can use to train their models
OpenAI also pointed out that any data that fine-tunes GPT-3.5 turbo will be under the user and other companies cannot use the data to train other GPT models.
The company recommends mixing fine-tuning with techniques like information retrieval, function calling and prompt engineering to enjoy the best results.
Photo Credit: Open AI
