CTRL: A Conditional Transformer Language Model for Controllable Generation

TL;DR

CTRL is a large conditional transformer language model that enables explicit control over generated text's style, content, and task-specific attributes through control codes derived from natural data structures.

Contribution

The paper introduces CTRL, a 1.63 billion-parameter language model that incorporates control codes for targeted text generation and data source attribution, advancing controllability in language models.

Findings

CTRL effectively controls style, content, and task-specific attributes.

The model can predict data source likelihoods, aiding data analysis.

Multiple pretrained versions of CTRL are publicly available.

Abstract

Large-scale language models show promising text generation capabilities, but users cannot easily control particular aspects of the generated text. We release CTRL, a 1.63 billion-parameter conditional transformer language model, trained to condition on control codes that govern style, content, and task-specific behavior. Control codes were derived from structure that naturally co-occurs with raw text, preserving the advantages of unsupervised learning while providing more explicit control over text generation. These codes also allow CTRL to predict which parts of the training data are most likely given a sequence. This provides a potential method for analyzing large amounts of data via model-based source attribution. We have released multiple full-sized, pretrained versions of CTRL at https://github.com/salesforce/ctrl.