Models

← All models

InstructGPT 6B

OpenAI · 2022-01-27

InstructGPT 6B is OpenAI's 6B-scale PPO-ptx policy from the InstructGPT work announced in January 2022. It was initialized from GPT-3, fine-tuned with supervised learning on human demonstrations, and further trained with RLHF via PPO while mixing pretraining-distribution gradients to reduce performance regressions.

Benchmark scores