# CTRL: A Conditional Transformer Language Model for Controllable Generation

**Authors:** Nitish Shirish Keskar, Bryan McCann, Lav R. Varshney, Caiming Xiong, Richard Socher  
**arXiv:** [1909.05858](https://arxiv.org/abs/1909.05858)  
**Date:** 2019-09-11  
**Summary page:** https://www.summarizepaper.com/en/paper/1909.05858v1/

## Key points

- Authors introduce CTRL, a 1.6 billion-parameter conditional transformer language model
- CTRL allows users to control style, content, and task-specific behavior through control codes derived from natural co-occurrence with raw text
- Users can influence text generation outcomes while benefiting from unsupervised learning
- CTRL's ability to predict likelihood of specific parts of training data enables analyzing vast amounts of data through model-based source attribution
- Multiple full-sized pretrained versions of CTRL are freely available on github.com/salesforce/ctrl

## Summary

, , , , 

In their paper titled "CTRL: A Conditional Transformer Language Model for Controllable Generation," authors Nitish Shirish Keskar, Bryan McCann, Lav R. Varshney, Caiming Xiong, and Richard Socher introduce CTRL, a cutting-edge 1.6 billion-parameter conditional transformer language model designed to address the limitations of existing large-scale models. CTRL allows users to control various aspects of text generation such as style, content, and task-specific behavior through the use of control codes derived from natural co-occurrence with raw text. This unique feature enables users to influence text generation outcomes while retaining the benefits of unsupervised learning. Additionally, CTRL's ability to predict the likelihood of specific parts of training data opens up possibilities for analyzing vast amounts of data through model-based source attribution. To further facilitate research and application development in this area, multiple full-sized pretrained versions of CTRL are freely available on github.com/salesforce/ctrl. This release marks a significant advancement in natural language processing and provides researchers and practitioners with a powerful tool for generating controlled and tailored text outputs across various domains and applications.