{
 "slug": "InstructGPT",
 "title": "InstructGPT",
 "type": "model",
 "short_desc": "OpenAI's 2022 RLHF-aligned GPT-3 variants; demonstrated that alignment training beats scale on human preference and set the template for ChatGPT.",
 "categories": [
  "OpenAI models",
  "Models",
  "2022 model releases",
  "Alignment and safety"
 ],
 "infobox": {
  "Developer": "OpenAI",
  "Family": "GPT",
  "Predecessor": "GPT-3",
  "Successor": "GPT-3.5 series (ChatGPT)",
  "Announced": "January 2022 (paper March 2022)",
  "Architecture": "Decoder-only Transformer (GPT-3 bases)",
  "Parameters": "1.3B, 6B, and 175B variants",
  "Post-training": "SFT + reward model + PPO (RLHF)",
  "License / access": "Proprietary; served as the default GPT-3 API models from 2022",
  "Paper/report": "arXiv:2203.02155"
 },
 "infobox_links": {
  "Developer": [
   "OpenAI"
  ],
  "Family": [
   "GPT_(model_family)"
  ],
  "Predecessor": [
   "GPT-3"
  ],
  "Successor": [
   "GPT-3.5"
  ],
  "Architecture": [
   "Transformer_(architecture)",
   "GPT-3"
  ],
  "Post-training": [
   "Supervised_fine-tuning",
   "Reinforcement_learning_from_human_feedback"
  ],
  "License / access": [
   "GPT-3"
  ]
 },
 "aliases": [],
 "family": "GPT (model family)",
 "org": "OpenAI",
 "registry_status": "stub",
 "words": 291,
 "references": 3,
 "outbound": [
  "Anthropic",
  "Constitutional_AI",
  "Direct_preference_optimization",
  "GPT-3",
  "GPT-3.5",
  "GPT-4",
  "GPT_(model_family)",
  "Hugging_Face",
  "Llama_(model_family)",
  "OpenAI",
  "Reinforcement_learning_from_human_feedback",
  "Reward_model",
  "Supervised_fine-tuning",
  "Transformer_(architecture)"
 ],
 "inbound": [
  "Andrej_Karpathy",
  "Flan-T5",
  "GPT-3",
  "GPT-3.5",
  "GPT-4_Turbo",
  "GPT_(model_family)",
  "Ilya_Sutskever",
  "Instruction_tuning",
  "Large_language_model",
  "Pretraining",
  "Reinforcement_learning_from_human_feedback",
  "Reward_model",
  "Richard_Sutton",
  "Supervised_fine-tuning"
 ],
 "url": "wiki/InstructGPT.html",
 "built": "2026-07-24"
}
