Dataset Card for OmniaBench

• Paper: OmniaBench: Benchmarking General AI Agents Across Diverse Scenarios (coming soon)
• Homepage: https://scuuy.github.io/OmniaBench/
• Repository: https://github.com/scuuy/OmniaBench
• Point of Contact: see the paper / GitHub repository

Dataset Summary

OmniaBench is a broad, diagnostic benchmark for evaluating general AI agents. Scenario knowledge is distilled from app stores, product documents, industry resources, and web retrieval into a hierarchical taxonomy spanning ToC, ToB, and ToE with 90 level-1 and 354 level-2 domains. On top of this taxonomy, executable environments are built and tasks are synthesized through four complementary construction routes, then scored along a ten-dimensional capability taxonomy and eight atomic difficulty factors.

This repository hosts the 644-task challenging subset, released for cost-efficient, contamination-resistant leaderboard evaluation, as four route files (route1.json–route4.json). It is the companion dataset to the OmniaBench GitHub repository, which provides the evaluation harness that consumes these files directly.

Dataset Structure

Each route file is a JSON array of task items. The four routes correspond to four different task construction methods and are evaluated with different scorers:

Route file	Route name	Tasks	Turn mode	Scorer
route1.json	DAG (main body)	354	multi-turn	rubric
route2.json	Solver	60	single-turn	rubric
route3.json	Program	30	single-turn	verifier (executable check code)
route4.json	DAG-S	200	single-turn	rubric

Data Fields

Fields are route-specific since each construction route captures different information. Common fields across most routes:

• global_id / task_id / env_id: task and environment identifiers
• task: the natural-language task instruction given to the agent
• init_config: initial state configuration for the task's environment
• rubrics: the weighted rubric checklist (or, for route3, see verifier_code below) used to score the agent's trajectory

Route-specific fields:

• route1 / route2: env_item (full environment definition: tools, class code, constraints, introduction), persona / persona_card_md / persona_idx (user-simulator persona for the multi-turn conversation), env_class_name, lang
• route1: also includes fs_inputs / fs_runtime_bootstrap for tasks requiring a filesystem sandbox (see runtime_assets/fs_bundle/ in the GitHub repo)
• route3: candidate_tools, env_class_code, constraints_rules, environment_introduction, ground_truth_answer, verifier_code (executable Python check function used for binary pass/fail scoring instead of rubric judging)
• route4: candidate_tools, env_class_code, env_class_name, persona / persona_card_md / persona_idx, lang

Some internal bookkeeping fields used only during dataset construction (e.g. domain taxonomy labels, construction-pipeline debug metadata) have been stripped from this release and are not required by the evaluation harness.

Data Splits

This repository contains a single, fixed 644-task challenging subset (not split into train/val/test): 354 + 60 + 30 + 200 tasks across the four routes above. This split is used as-is for leaderboard evaluation; see the paper for details on how it was selected from the full 1,431-task collection.

Usage

See evaluation/README.md in the GitHub repository for how to download this dataset and run model evaluation against it with the provided evaluation harness.

Licensing Information

The dataset license is still being finalized and will be published alongside the arXiv paper. Until then, treat this dataset as research-use only; do not redistribute. Check back here or the GitHub repository for the finalized terms.

Citation

@misc{shen2026omniabenchbenchmarkinggeneralai,
      title={OmniaBench: Benchmarking General AI Agents Across Diverse Scenarios}, 
      author={Chengyu Shen and Yujie Fu and Gangtao Xin and Yanheng Hou and Wenlong Fei and Guojie Zhu and Jiawei Li and Hongcheng Gao and Runming He and Zhen Hao Wong and Meiyi Qiang and Hao Liang and Zhao Cao and Hao Jiang and Chong Chen and Wentao Zhang},
      year={2026},
      eprint={2607.14989},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.14989}, 
}