CrowdMath: A Dataset of Crowdsourced Mathematical Research Discussions

In three linesCrowdMath is a dataset of 164 annotated discussion chains from the MIT PRIMES–Art of Problem Solving program (2016-2025), capturing collaborative open-problem solving. Posts are labeled by functional role (partial progress, error, repair). Six frontier models achieve 83-88% accuracy on next-post prediction but only 0.42 macro-F1 on post-role classification.

Summary generated by Claude — human-verified