Research OpenAI LLM Proprietary

OpenAI InstructGPT

RLHF alignment applied to GPT-3 — the precursor to ChatGPT

Released 2022-01-27 · ~175B

Overview

GPT-3 fine-tuned to follow instructions using human feedback

Specifications

DeveloperOpenAI
Release date2022-01-27
Model typeLLM
Parameters~175B
ArchitectureTransformer (GPT-3 + RLHF)
LicenseProprietary
Input modalitiestext
Output modalitiestext

Benchmarks

Benchmark InstructGPT
TruthfulQA 58.0%
RealToxicityPrompts 0.3%
Helpfulness 85.0%

Availability

API Only OpenAI API

Official references