InstructGPT

Retired · Language

InstructGPT is a set of GPT-3 language models that OpenAI fine-tuned with human feedback to follow user instructions, offered through the OpenAI API, where they became the default language models in January 2022. The research models come in 1.3B, 6B and 175B sizes; OpenAI says the API versions used a slightly different fine-tuning procedure. [1] [2] Primary source

Timeline of InstructGPT →

Claims and evidence

  • Released Date on which OpenAI announced InstructGPT and made it the default language models on its API. The post says these models had been in beta on the API for more than a year, and that an earlier version was deployed in January 2021 (footnote A). Primary source[1]page date; first and fourth paragraphs
  • Deprecated Primary source[3]2023-07-06: GPT and embeddings
  • Retired Primary source[3]2023-07-06: GPT and embeddings, InstructGPT models
  • Status Retired Primary source[3]2023-07-06: GPT and embeddings, InstructGPT models
  • Replaced by GPT-3.5 TurboThe deprecations page names gpt-3.5-turbo-instruct as the recommended replacement for text-ada-001, text-babbage-001, text-curie-001 and text-davinci-001. Primary source[3]2023-07-06: GPT and embeddings, InstructGPT models
  • Derived from (fine tune) GPT-3Fine-tuned from GPT-3 with supervised learning on human demonstrations and reinforcement learning from human feedback. Primary source[1]Methods [2]Model Details [4]
  • Change · Other Trained with labeler demonstrations, a reward model and PPO to do the task the user asks for, instead of only predicting the next word of internet text.Compared with GPT-3 Primary source[1]
  • Change · Safety OpenAI says it makes up facts less often than GPT-3 and shows small decreases in toxic output.Compared with GPT-3 Primary source[1]
  • Change · Availability Replaced the GPT-3 base models as the default language models on the OpenAI API.Compared with GPT-3 Primary source[1]
  • Parameters 1.3B, 6B, and 175BThe three sizes in the paper and model card; value is the largest. OpenAI says the versions deployed on the API were fine-tuned with a slightly different procedure, and the model index lists no close API match for the paper models apart from davinci-instruct-beta (175B SFT). Primary source[2]Model Details
  • Access API Primary source[1]first paragraph [2]Model Use

Lineage

Predecessors

No known predecessor.

Successors

No known successor.

Based on

  • GPT-3 · 11 June 2020 · derived (fine tune)

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

  • GPT · 11 June 2018
  • GPT-2 · 14 February 2019
  • GPT-3 · 11 June 2020

All descendants

None.

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from OpenAI on AI Radar →