ctrl

1. Model Details
2. Uses
3. Bias, Risks, and Limitations
4. Training
5. Evaluation
6. Environmental Impact
7. Technical Specifications
8. Citation
9. Model Card Authors
10. How To Get Started With the Model

Model Details

Model Description

The CTRL model was proposed in CTRL: A Conditional Transformer Language Model for Controllable Generation by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher. It's a causal (unidirectional) transformer pre-trained using language modeling on a very large corpus of ~140 GB of text data with the first token reserved as a control code (such as Links, Books, Wikipedia etc.). The model developers released a model card for CTRL, available here.

In their model card, the developers write:

The CTRL Language Model analyzed in this card generates text conditioned on control codes that specify domain, style, topics, dates, entities, relationships between entities, plot points, and task-related behavior.

• Developed by: See associated paper from Salesforce Research
• Model type: Transformer-based language model
• Language(s) (NLP): Primarily English, some German, Spanish, French
• License: BSD 3-Clause; also see Code of Conduct
• Related Models: More information needed
  • Parent Model: More information needed
• Resources for more information:
  • Associated paper
  • GitHub repo
  • Developer Model Card
  • Blog post

Uses

Direct Use

The model is a language model. The model can be used for text generation.

Downstream Use

In their model card, the developers write that the primary intended users are general audiences and NLP Researchers, and that the primary intended uses are:

1. Generating artificial text in collaboration with a human, including but not limited to:

• Creative writing
• Automating repetitive writing tasks
• Formatting specific text types
• Creating contextualized marketing materials

2. Improvement of other NLP applications through fine-tuning (on another task or other data, e.g. fine-tuning CTRL to learn new kinds of language like product descriptions)
3. Enhancement in the field of natural language understanding to push towards a better understanding of artificial text generation, including how to detect it and work toward control, understanding, and potentially combating potentially negative consequences of such models.

Out-of-Scope Use

In their model card, the developers write:

• CTRL should not be used for generating artificial text without collaboration with a human.
• It should not be used to make normative or prescriptive claims.
• This software should not be used to promote or profit from:
  • violence, hate, and division;
  • environmental destruction;
  • abuse of human rights; or
  • the destruction of people's physical and mental health.

Bias, Risks, and Limitations

Significant research has explored bias and fairness issues with language models (see, e.g., Sheng et al. (2021) and Bender et al. (2021)). Predictions generated by the model may include disturbing and harmful stereotypes across protected classes; identity characteristics; and sensitive, social, and occupational groups.

In their model card, the developers write:

We recognize the potential for misuse or abuse, including use by bad actors who could manipulate the system to act maliciously and generate text to influence decision-making in political, economic, and social settings. False attribution could also harm individuals, organizations, or other entities. To address these concerns, the model was evaluated internally as well as externally by third parties, including the Partnership on AI, prior to release.

To mitigate potential misuse to the extent possible, we stripped out all detectable training data from undesirable sources. We then redteamed the model and found that negative utterances were often placed in contexts that made them identifiable as such. For example, when using the ‘News’ control code, hate speech could be embedded as part of an apology (e.g. “the politician apologized for saying [insert hateful statement]”), implying that this type of speech was negative. By pre-selecting the available control codes (omitting, for example, Instagram and Twitter from the available domains), we are able to limit the potential for misuse.

In releasing our model, we hope to put it into the hands of researchers and prosocial actors so that they can work to control, understand, and potentially combat the negative consequences of such models. We hope that research into detecting fake news and model-generated content of all kinds will be pushed forward by CTRL. It is our belief that these models should become a common tool so researchers can design methods to guard against malicious use and so the public becomes familiar with their existence and patterns of behavi