Every game is a data-collection task in disguise. Flip any card to see exactly what it collects and why a client wants it.
These are formats we've already built. Drop your own data into one - or we design a brand-new game around your task. Hover any game to see what it collects.
Written Arabic usually drops short vowels - so the same letters can be many different words. Restoring diacritics (tashkeel) is one of the hardest, highest-value problems in Arabic NLP.
Arabic OCR - especially handwriting and historical manuscripts - is notoriously unreliable. Human verification turns raw OCR into gold transcription data.
Rating answer quality is the backbone of RLHF and model evaluation. A balanced crowd of native speakers produces the preference data that aligns models.
Arabic is dozens of distinct spoken dialects, not one language. Labeled dialect data is scarce - and essential for models that understand how people actually talk.
Safety classifiers need culturally-aware judgments of what counts as toxic or offensive - context that off-the-shelf models routinely get wrong.
Paraphrase and semantic-similarity labels train search, retrieval, and deduplication - and good similarity data is hard to come by.
Entity tagging powers everything from knowledge graphs to redaction. NER suffers from sparse, inconsistent labels - exactly what a big crowd fixes.
True/false judgments at speed build fact-checking and claim-verification datasets - and keep players in flow while they label.
Fill-in-the-blank responses generate completion and next-word data, and reveal how real speakers phrase things in context.
Fast multiple-choice rounds collect classification labels across any domain you can phrase as a question.
A second MCQ format keeps the same labeling task fresh - variety is what keeps the crowd annotating for hours.
Endless streak play maximizes labels-per-session and naturally surfaces the questions players find hardest.
