Large-scale language models show promising text generation capabilities, but users cannot control their generated content, style or train them for multiple supervised language generation tasks.