About the job
We are on a mission to build machines that understand the world and make them safely accessible to all. Data quality is foundational to this process. Machines (or Large Language Models to be exact) learn in similar ways to humans: by way of feedback. By labeling, ranking, auditing, prompting, red teaming, and correcting output, you will improve Large Language Model’s performance for iterations to come, thus having a lasting impact on Cohere’s tech.
Responsibilities
- Evaluate and improve model safety: Label, rank, audit, and refine human- and model-generated text to improve safety, quality, and policy alignment, including content that may be sexual, violent, or psychologically disturbing.
- Apply nuanced safety judgment: Assess model outputs against detailed safety guidelines, rubrics, and style standards, making consistent decisions across ambiguous, sensitive, and context-dependent cases.
- Create prompts and safety test cases: Write realistic prompts, user scenarios, and adversarial examples that help evaluate model behavior across safety categories and uncover unsafe, evasive, over-refusing, or policy-inconsistent responses.
- Support quality and calibration: Identify annotation inconsistencies or unclear guidelines, and provide actionable feedback on recurring edge cases, model failures, and opportunities to improve data quality.
- Work with precision and independence: Complete annotation tasks with strong attention to detail, while being comfortable working independently in a globally distributed, asynchronous team environment.
Qualifications
Minimum
- 1+ years of experience in Content Moderation, Trust and Safety, AI data annotation, LLM evaluation, or a related analytical role, with exposure to quality assurance, red teaming, and/or prompt engineering preferred.
- Experience applying detailed guidelines to complex and sensitive content, with strong contextual and sociocultural judgment and the ability to recognize and manage personal bias.
- Emotional resilience: Comfort working with content that contains unsafe, explicit, and/or toxic content, including content of a sexual, violent, or psychologically disturbing nature.
- Excellent command of written English and the ability to clearly justify content evaluations, including why an output is safe, unsafe, high-quality or low-quality.
- Strong attention to detail and commitment to accuracy, with the ability to maintain consistency across high-volume and monotonous tasks.
- Strong execution in a remote environment, including good time management, comfort using new tools, and the ability to work independently in a global, asynchronous team.
Preferred
- Bonus points if you are fluent in another language!