Synthetic data is most useful for agent training when its underlying truth is known, yet the solver cannot simply read that truth from the files. These tasks are built around that distinction: each environment contains observable data, while the mechanisms and records used to generate and verify it are withheld. The result is a way to train analytical reasoning against grounded targets rather than unverified answers.
AI Dataset2026
Zhanli-Li/DataWorld-Train-2k-part2
Provides self-contained synthetic environments for training data-analysis agents. This slice contains 1,538 tasks, each graded with a rubric designed to evaluate analytical findings.
Introduction
Information
- Websitehuggingface.co
- OrganizationsLongcat Interaction Team, Meituan, Nanjing University
- AuthorsZhanli Li, Zeyuan Xu, Chaoxu Pang, Zeyuan Liu, Anyang Ji, Te Sun, Kongcheng Zhang, Wenbo Xu, Xubin Yue, Shipeng Luo …
- Published date2026/10/16
Categories
More Items
AI Dataset2026
Evaluates data agents on 100 tasks that require evidence-backed conclusions from raw files spanning multiple business systems. Each task comes from a synthetic world and tests either causal explanations or downstream impact analysis against a task-specific rubric.
AI Dataset2026
Diagnoses physical plausibility in generated videos by scoring adherence to 13 observable laws. It pairs benchmark prompts and model outputs with quality-controlled human ratings for fine-grained evaluation.
AI Dataset2026
Provides continuous human whole-body manipulation demonstrations for robot policy training, including locomotion and coordinated two-hand actions. Motion is reconstructed from egocentric observations, with reported whole-body joint error of about 3 cm.