DSGym: A holistic framework for evaluating and training data science agents

It integrates existing benchmarks and adds novel scientific analysis tasks (DSBio, 90 bioinformatics tasks) and Kaggle competitions (DSPredict, 92 competitions).

Using DSGym, the team trained a 4B model (Qwen3-4B-DSGym-SFT-2k) on 2,000 synthetic query-trajectory pairs, achieving state-of-the-art performance among open-source models on general data science benchmarks.