RLHF
Reinforcement learning from human feedback. People compare or rate model answers, and those judgments are used to train the model to prefer better responses. Many rating and ranking jobs feed this process.
Rubric
A written checklist for judging an answer: accuracy, safety, completeness, tone. Expert roles often ask you to apply a rubric or to write one for your field.
Red-teaming
Trying to make a model misbehave on purpose, safely and under instructions, so its weaknesses can be fixed. Safety and policy roles use this term.
Prompt and response writing
Writing example questions and ideal answers that show a model what good work in your field looks like. Often called SFT data, for supervised fine-tuning.
Annotation and labelling
Tagging text, images, audio or video so a model can learn from it. Usually paid per hour or per task and open to generalists.
Evaluation
Testing a model’s answers against a standard and recording what went wrong. Coding, maths and medical evaluation roles are among the best paid.
Talent pool
A list of people a platform may contact for future projects. Joining one is not the same as being offered paid work.
Per task versus per hour
Some platforms pay a fixed amount per task. You can only compare that with an hourly rate once you know how long a task really takes, including review and revisions.
Sources and further reading
Platform requirements change. Always confirm them on the official listing.