benchmarking

Data science agents must inspect datasets, design workflows, run code, and return verifiable answers, not just autocomplete Pandas code. DSGym, introduced by researchers at Stanford University, Together AI, Duke University, …