CTRL: A Conditional Transformer Language Model for Controllable Generation

TLDR

Large‑scale language models excel at text generation, yet users lack fine‑grained control over generated content. The authors 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. CTRL is trained on massive text corpora with control codes derived from naturally co‑occurring structure, providing explicit control while preserving the benefits of unsupervised learning. The control codes allow CTRL to predict the most likely training data segments for a given sequence, enabling model‑based source attribution, and several full‑size pretrained versions have been made 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.