Reinforcement Learning with Human Feedback (RLHF)

TRL

RLHF on GPT-2

Training the model to generate more positive-sentiment text using PPO

Colab Notebook — RLHF Positive Sentiment with PPO on GPT-2colab.research.google.com

Training the model to generate positive, neutral, or negative text (Controlled Sentiment) by specifying a prefix in the input

Colab Notebook — RLHF Controlled Sentiment on GPT-2colab.research.google.com