Purposes utilizing the secure mannequin names for base GPT-3 fashions (ada, babbage, curie, davinci) will routinely be upgraded to the brand new fashions listed above on January 4, 2024. The brand new fashions may also be accessible within the coming weeks for early testing by specifying the next mannequin names in API calls: ada-002, babbage-002, curie-002, davinci-002.
Builders utilizing different older completion fashions (reminiscent of text-davinci-003) might want to manually improve their integration by January 4, 2024 by specifying gpt-3.5-turbo-instruct within the “mannequin” parameter of their API requests. gpt-3.5-turbo-instruct is an InstructGPT-style mannequin, educated equally to text-davinci-003. This new mannequin is a drop-in alternative within the Completions API and will likely be out there within the coming weeks for early testing.
Builders wishing to proceed utilizing their fine-tuned fashions past January 4, 2024 might want to fine-tune replacements atop the brand new base GPT-3 fashions (ada-002, babbage-002, curie-002, davinci-002), or newer fashions (gpt-3.5-turbo, gpt-4). As soon as this function is offered later this yr, we are going to give precedence entry to GPT-3.5 Turbo and GPT-4 fine-tuning to customers who beforehand fine-tuned older fashions. We acknowledge that migrating off of fashions which can be fine-tuned by yourself knowledge is difficult. We will likely be offering assist to customers who beforehand fine-tuned fashions to make this transition as clean as potential.
Within the coming weeks, we are going to attain out to builders who’ve just lately used these older fashions, and can present extra data as soon as the brand new completion fashions are prepared for early testing.
