text-davinci-002

text-davinci-002 is an OpenAI language model belonging to the GPT-3.5 series. It is an instruction-tuned variant from the GPT-3/InstructGPT family, built on top of code-davinci-002 and trained with supervised fine-tuning (the so-called FeedME method) rather than reinforcement learning from human feedback (RLHF).

The model was released on 15 March 2022 alongside code-davinci-002, introducing text edit and insert capabilities. It supported a 4,097-token context window, and its training data extended up to June 2021.

text-davinci-002 became well known in research on reasoning in language models — Kojima et al. (2022) used it to demonstrate the Zero-shot Chain-of-Thought technique, raising accuracy on the GSM8K benchmark from 10.4% to 40.7% simply by appending the phrase "Let's think step by step" to the prompt.

OpenAI announced the retirement of older GPT-3 and GPT-3.5 models on 6 July 2023; text-davinci-002 was finally shut down on 4 January 2024. The recommended successor is gpt-3.5-turbo-instruct.