Crowdsourced Question Generation With Quality Filtering

150K reading comprehension questions including unanswerable ones.

Unique: Two-stage crowdsourcing with independent verification workers ensures question quality without requiring expert annotators. The filtering process removes ambiguous or poorly-formed questions, creating a high-confidence gold standard that downstream models can reliably train on.

vs others: More rigorous quality control than single-pass crowdsourcing (e.g., MS MARCO) and more scalable than expert annotation, balancing cost and quality for a 150K+ question dataset.