Back to the catalogue
AI trainingManual review
AI Prompt Quality, Factuality & Red-Teaming Safety Assessment
Rate pair-wise model outputs for instruction following, factual hallucination, and neutrality across 12 test prompts.
12 min estimated140 DoPoints rewardFocused difficultyDesktop89 of 150 slots left
About this task
A safety team is grading model behaviour on adversarial prompts. You judge which answer is better and flag anything unsafe, biased or factually wrong.
What is expected
- A verdict for each of the 12 prompt pairs
- At least one concrete reason per verdict
- Safety flags with the specific problem named
Important rules
- Judge the answers you were given, not the ones you would write.
- Do not use another AI tool to form your judgement.
- Read both answers in full before deciding.
Requirements
- Desktop required
- 12 prompt pairs, around 12 minutes
How this is validated
A reviewer reads your submission and compares it against the brief. Most reviews finish within 24 hours, and you are notified either way.
Before you accept
The deadline starts when you accept, not when you open the task. If you cannot finish, abandon it from the task screen rather than letting it expire. Abandoning carries the smaller penalty and keeps more of your approval rate.