PPLM

Why we included this project

PPLM is the reference implementation of a widely cited approach to controlled text generation. Instead of retraining a language model, you steer its output by plugging in one or more small attribute models that push generation toward a chosen topic, sentiment, or other property. The repo ships working code for topic steering, sentiment control, and detoxification on GPT-2, along with the discriminator models and hyperparameters from the paper. Because the base model is used as-is, it offers a practical way to study how to influence generation without the hardware that fine-tuning a large model would demand. It also integrates with Hugging Face Transformers, so it is easy to try in an existing setup.