DataPraxis is an analytics‑task synthesis framework built on verified workspace evolution. The framework first enriches a workspace step by step with validated findings and artifacts, checking the analytical result at each step. It then inherits and combines existing workspaces to surface deeper questions. Tasks are constructed from the accumulated findings and their dependencies, with verification performed both during workspace evolution and after task assembly.
Using this pipeline we created the DataPraxis-5K dataset, comprising 5,000 verified trajectories across diverse domains and output formats. Models trained on the dataset, such as DataPraxis-27B, achieve 55.11 on DABStep and 51.9 on JobBench, matching or surpassing frontier models like GLM‑5.2.
DataPraxis-5K is publicly released on HuggingFace, enabling further research on training professional analytics agents.
Review