What the work is
RLHF tasks: ranking, rating, prompt writing and red-teaming
Across the platforms we read, the task named most is judging a model's output; writing for the model comes second, and red-teaming is rare.
Email me RLHF task changes Say which RLHF task suits you and your field. We email you when a platform on the table names new tasks, tracks or entry tests for it, and only then.
Judging model answers
6 of the 12 platforms describe ranking, comparing or rating responses. DataAnnotation says its projects commonly include reviewing AI-generated text, ranking responses and checking facts, and Prolific invites participants to evaluate machine learning models.
Source: RLHF task types board
Writing prompts, answers and rubrics
3 platforms name writing work. Outlier asks you to think of a difficult question in your field of study and then write the correct answer, and Mindrift asks for realistic prompts, the kind a real user would actually ask AI.
Source: RLHF task types board
Probing models for weaknesses
Red-teaming appears on 1 platform's pages: Alignerr's AI Red Team Tester role, described as probing cutting-edge models for weaknesses to make them safer.
Source: RLHF task types board