OpenAI Brings Fine-Tuning to GPT-3.5 Turbo and GPT-4
OpenAI has launched fine-tuning capabilities for its GPT-3.5 Turbo large language model API, with plans to do the same with GPT-4 this fall. The fine-tuning will allow developers to customize the generative AI model to perform better on specific use cases.
Turbo serves as a midway point between GPT-3.5 and GPT-4 and has been the model fueling ChatGPT, Bing AI, and other marquee generative AI projects. OpenAI described the success of early tests with the fine-tuned versions of GPT-3.5 Turbo matching or even exceeding baseline GPT-4 on narrow tasks. Fine-tuning allows businesses to tweak the model for their particular needs. Private beta users have leveraged fine-tuning to boost capabilities like instruction following, output formatting, and tone. For example, developers can ensure the AI consistently responds in a certain language when prompted. The upgrade enables full customization for businesses and developers seeking specialized enhancements like tailored formatting, instruction following, and tone.
The fine-tuning costs $0.008 per 1,000 tokens. Once trained, the model costs $0.012 per thousand tokens for input and $0.016 for every thousand tokens of response. By comparison, the standard GPT-3.5 Turbo API costs $0.0015 and $0.002 per thousand tokens for input and response, respectively, for the 4,000 token context window and $0.003 and $0.004 for the 16,000 token context window. However, the overall price with fine-tuning can be lower, according to OpenAI, because the technique enables users to cut prompt length by up to 90% while maintaining performance and speeding up interactions.
“Since the release of GPT-3.5 Turbo, developers and businesses have asked for the ability to customize the model to create unique and differentiated experiences for their users. With this launch, developers can now run supervised fine-tuning to make this model perform better for their use cases,” OpenAI explained in a blog post. “Fine-tuning with GPT-3.5-Turbo can also handle 4k tokens—double our previous fine-tuned models.”
OpenAI only added GPT-3.5 Turbo and GPT-4 to its API menu in July, along with the Whisper speech-to-text and DALL-E 2 image generator models. The news accompanied the outline for a plan to phase out older models accessible model generally available through the API. The API improvement comes after OpenAI added several upgrades to ChatGPT to make it more responsive and personalized, particularly the Custom Instructions feature. The company also just made its first acquisition, buying digital products and design startup Global Illumination.