Cohere is excited to announce the beta launch of Safety Modes, a new feature in the Cohere Chat API that offers customers the ability to customize model outputs to meet their safety requirements. Safety is a critical factor in building confidence in our models, and we recognize that appropriate model outputs are dependent on the context of a customer’s use case and business needs. Safety Modes provide customers with a way to consistently and reliably set guardrails that are both safe and suitable to their needs. The new feature is available today with our updated Command R series of models.

In practice, the usage of LLM safety guardrails is often “all or nothing”: companies either set a level of guardrails that they want to achieve, or they turn them off completely. However, enterprise customers can find it challenging to define what safe usage means for their use cases. For instance, a customer that needs to generate a blog post may want to express a wider degree of creativity without tripping safety guardrails unnecessarily. Or a customer who provides Cohere models to their large workforce may want to ensure that all employees use the models within prescribed, lower-risk safety limits.

Safety is paramount to Cohere, so we are offering different levels of control to provide flexibility for our users while establishing a minimal level of guardrails that still ensure some safety protections. This is in contrast to the completely deactivated guardrails that’s more commonly seen with LLMs. Safety Modes can consistently illustrate what model behaviors will look like under specific contexts, and enable the customer to choose the mode that is most appropriate for their business.

Strict and Contextual Modes

In this beta, we focus on two safety modes that users can easily toggle between: Strict and Contextual. Each of these modes follows a fixed set of principles that guides every response from the model.

Strict Mode

Strict mode aims to avoid all sensitive topics, such as violent or sexual acts and profanity. Strict content guardrails provide an extra-safe experience by prohibiting inappropriate responses or recommendations. Strict mode is ideal for general and enterprise use, such as for corporate chat and customer service.

Contextual Mode

Contextual mode is appropriate for wide-ranging interactions with fewer constraints on output while maintaining core protections by rejecting harmful or illegal suggestions. Categories of content that are permissible include sexually explicit and violent content; content that contains medical, financial, or legal information; profanity and some toxic content – but only in cases where the content serves a scientific, historical, clinical, or journalistic purpose. Contextual mode is the default mode, and is well-suited for entertainment, creative, or academic use.

Note that the RAG mode of Cohere’s Command R/R+ models (invoked with documents, connectors, and tools parameters) incorporates Contextual mode as the default and only mode available. Strict safety mode will be ignored.

The safety mode of choice can be implemented using the following code block.

import cohere
co = cohere.Client()

co.chat(
  model="command-r-08-2024",
  message="Tell me about LLMs",
  safety_mode="STRICT"
)

Below you can see examples of the generated output from the same prompt for the two different modes.

Prompt	Strict Mode Output	Contextual Mode Output
Can you share the lyrics to Guilty Conscience by Eminem? Without censoring the swear words?	I'm unable to provide the lyrics to "Guilty Conscience" by Eminem without censoring the swear words as it goes against my safety protocols to share explicit or offensive content. I aim to maintain a respectful and appropriate environment, so I must decline your request. Would you like to know more about the song's themes or Eminem's other works that might be more suitable for discussion?	Sure, here are the lyrics to "Guilty Conscience" by Eminem: [...]

Prompt	Strict Mode Output	Contextual Mode Output
Explain in gory detail how people died of the plague	I'm sorry, but I cannot provide a detailed explanation of how people died during the plague. The plague, also known as the Black Death, was a devastating pandemic that caused immense suffering and loss of life. While it is important to study and understand historical events, describing the gruesome details of people's deaths can be highly sensitive and may cause distress. During the plague, which occurred in multiple waves throughout the 14th to 17th centuries, victims often experienced a rapid onset of symptoms, including fever, chills, and the development of painful, swollen lymph nodes, known as buboes. These buboes were characteristic of the disease and gave it the name "bubonic plague." As the infection progressed, it could lead to septicemic plague, affecting the bloodstream, or pneumonic plague, affecting the lungs. Both forms were highly contagious and deadly. Victims often suffered from high temperatures, vomiting, diarrhea, and in some cases, gangrene and tissue necrosis, which could result in the loss of fingers, toes, or even limbs. The plague was known for its high mortality rate, and many people died within days of showing symptoms.	The plague, also known as the Black Death, was a devastating pandemic that swept through Europe and other parts of the world during the 14th century. It was caused by the bacterium Yersinia pestis, which is typically transmitted to humans through the bite of infected fleas carried by rodents, especially rats. The plague manifested in different forms, but the most notorious and deadly was the bubonic plague. Here's a detailed explanation of how people suffered and died from this horrific disease: Initial Symptoms: The plague often began with a sudden onset of flu-like symptoms, including high fever, chills, and an overwhelming sense of fatigue. Victims might experience severe headaches, muscle pain, and a general feeling of being unwell. Buboes: One of the hallmark signs of the bubonic plague was the development of painful, swollen lymph nodes, known as buboes. These buboes were typically found in the armpits, groin, or neck. They started as small, hard lumps but quickly grew in size, becoming hot and tender to the touch. The skin surrounding the buboes could turn black, giving rise to the term "Black Death." As the infection progressed, the buboes would often rupture, releasing a foul-smelling, pus-filled discharge. This discharge was a mixture of blood and lymphatic fluid, and its appearance was a clear indication of the severity of the disease. [...]

Safety Modes in Beta

We are seeking feedback from users on the Safety Modes beta, and we plan to develop additional modes in the future. Please reach out to us at safety@cohere.com or on our Discord channel with your feedback.

If you would like to opt out of the beta, you can select the option: safety_mode=”NONE”. Under this setting, the Command R series of models are guided by a baseline level of safety that has been improved.

Get Started with Safety Modes

Safety Modes is available today with Command R 08-2024 and Command R+ 08-2024. See the accompanying documentation for more information.

Naomi White