A useful way to train an analytical agent is to generate not just files and questions, but also the hidden process that produced them. That makes it possible to check whether a reported conclusion is supported by the observed data, rather than relying only on plausible-sounding answers.
What Sets It Apart
- Tasks pair heterogeneous, imperfect data with questions about causes or consequences, exercising data processing as well as interpretation.
- Each task includes its own environment and rubric, so it can be run and evaluated independently.
- The full training collection has 2,135 tasks across 1,037 worlds; this repository contains 597 tasks from worlds
p000000–p000499. - Audited generative mechanisms provide a basis for checking whether target findings can be recovered from the files, while the task setup withholds the world specification and causal graph from the agent.
Great Fit If...
It suits researchers training or evaluating agents that analyze files with code and need verifiable supervision. Look elsewhere if you need real-world business data or a measure of performance on naturally occurring user requests: the environments and tasks are synthetic, and valid insights beyond the rubric may not receive credit. This is part one of a split dataset, and a small number of its tasks require extracting archived task directories before use.
