OpenAI has finally introduced fine-tuning for its popular GPT-3.5 Turbo model. The announcement is aimed at businesses and developers to create supervised products that excel at specific tasks. According to OpenAI, a fine-tuned version of the GPT-3.5 Turbo model can match and even outperform advanced models like the GPT-4. So to learn more about custom fine-tuning on GPT-3.5 Turbo, follow along.
Fine-tune GPT-3.5 Turbo on Custom Knowledge Base
Fine-tuning allows users to customize the model and create something that tailors to their needs. We already know that GPT-3.5 Turbo is very fast at drawing inferences so in terms of performance, it was already far better than the competition. Now, with fine-tuning support, developers can leverage this ability to create new experiences, be it in the form of AI chatbots, documentation engines, AI assistants, coding assistants, etc. You can basically create a custom application for your business in your desired tone.
Keep in mind, OpenAI is keeping GPT-3 series models including babbage-002 and davinci-002 for fine-tuning purposes. However, the API endpoint has been changed and you can learn more about the updated endpoint from here. Besides that, the key feature of fine-tuning for GPT-3.5 Turbo is that all the training data are scrutinized using OpenAI’s Moderation API and GPT-4 powered moderation system. The company says it’s done to “detect unsafe training data that conflict with our safety standards.“
However, OpenAI has also made it clear that none of the private training material will be used to train OpenAI models. Apart from that, currently, GPT-3.5 Turbo can handle only 4k tokens at once. The company says that 16k context length and function calling support will be coming later this fall.
As for pricing, fine-tuning the GPT-3.5 Turbo model would cost $0.0080 per 1000 tokens for training; $0.012 per 1000 tokens for usage input; and $0.016 per 1000 tokens for usage output. Of course, it’s a bit pricier than Davinci and Babbage models, but you also get much better results from the fine-tuned version of GPT-3.5 Turbo.
