OpenAI opens up GPT-3.5 Turbo for customization
Developers will be able to train ChatGPT to focus on specific tasks, such as customizing the code, or to match the tone agreed upon for a particular company.
OpenAI has enabled companies to customize GPT-3.5 Turbo using their own data for training, allowing chatbots to be “more targeted,” meaning they can be designed for more specific tasks. For example, to respond to customers in the same style or with more concise wording.
Until now, customization was limited to GPT-3 variants, such as davinci-002 or bebbage-002.
The GPT-3.5 Turbo is a family of models that OpenAI introduced earlier this year, which it says is ideal for non-chatbot use. The model can process 4000 tokens simultaneously, which is twice as much as previous versions. The company added that the first testers used 90% shorter prompts after GPT-3.5 was prepared with fine-tuned instructions.
The price of GPT-3.5 is $0.0080 for 1000 training tokens, $0.0120 for 1000 incoming tokens, and $0.0120 USD for 1000 outgoing tokens.
The model is pre-trained, like GPT-4, on data up to September 2021, which can be supplemented with company data. OpenAI says that the information will not be used to train models outside the client company.
The company adds that the additionally customized GPT-3.5 Turbo model will be able to exceed the capabilities of the GPT-4 in some ways. The latter (which can read images in addition to text) will be opened for individual customization this fall.
After the release of GPT-4 in early March, OpenAI announced that it would be phasing in the update of existing models instead of creating new ones from scratch. At a recent Economic Times conference, the company’s CEO Sam Altman confirmed that OpenAI is not engaged in GPT-5 training, as a lot of preparatory work needs to be done before the new language model is launched.
