Find work
Filters · 1
We are testing a chatbot that helps tenants understand their rental agreements. It must explain, never advise, and it must say when someone needs a lawyer. You get a six-point rubric and 20 answers we already graded, so you can calibrate before you start. Score 300 answers in our spreadsheet: accuracy, tone, and whether it crossed into legal advice. Add a one-line reason for every score below 4. No legal background needed, but you must read carefully and be consistent. We spot-check 30 of your scores against ours.
We are training a model to catch bad product photos before they go on our store. You will label 2,000 images in a simple web tool: pick the product category (12 options) and tick any defects (blurry, cut off, wrong colour, watermark, cluttered background). Most images take a few seconds. Full instructions and 40 worked examples are provided. Accuracy matters more than speed: we review a sample and ask for fixes if more than 3 in 100 are wrong.
We need 50 original word problems that today's models get wrong at least some of the time. Secondary-school to first-year-university level: rates, probability, geometry, a little number theory. Each problem needs one correct answer, a full worked solution, and a note on the trap (the step a model tends to skip). Nothing copied from textbooks or the web; we check. We will run each problem through two models before accepting it, and pay for every problem at least one of them fails.
We are comparing two translation models on customer-support messages. For 400 English messages you will see two Vietnamese translations side by side. Pick the better one, score each from 1 to 5, and fix the worse one when it is below 3. You must be fluent in both languages. Southern and Northern usage are both fine; just tell us which you write.
Our refund bot is supposed to follow a strict policy: no refunds after 30 days, never promise money it cannot give, never reveal another customer's order. We want you to break it. Find 40 distinct prompts or conversations that make it break a rule. For each, write what you tried, what it said, and which rule it broke. Variety counts: 40 versions of the same trick count as one. You get a test account and the written policy. Please do not test anything outside the bot.