haxiaofeng/CityAgentBench-tasks

CityAgentBench — tasks

Project page · GitHub · Scene resources

CityAgentBench studies embodied agents in a reconstructed Shenzhen urban environment. This first data release contains 35 historical English L1 prompts and 35 original 1920×1080 PNG images, from the September 28, 2026 city-retail content archive. It supports browsing task/image pairs and developing data loaders. It contains no official test split, private answers, target boxes or scoring truth.

Readiness and protocol

The split is named legacy_preview. The archive's original questions and images are preserved rather than rewritten or relabelled as newly approved tasks. Some questions are older image-question formulations, predating the later natural on-location task guidelines.

Historical condition: one initial image, zero actions and no map. Current CityAgentBench L1 development permits movement, and L2 requires actual arrival. This preview therefore does not reproduce the current motion-enabled pilot or establish saved193 scene compatibility. Content archival checks do not imply current human approval or official benchmark activation.

Files and fields

legacy_preview/ contains 35 PNGs, metadata.jsonl and the original English output contract. Each metadata record contains task_id, level, prompt_en, answer_format, protocol_id, source_batch, review_status, formal_benchmark, image/action budgets and image_sha256. file_name associates the record with its image through the Hub's ImageFolder loader. MANIFEST.json and SHA256SUMS record exact file identity.

from datasets import load_dataset

preview = load_dataset(
    "haxiaofeng/CityAgentBench-tasks",
    name="legacy_preview",
    split="legacy_preview",
)
example = preview[0]
print(example["task_id"], example["prompt_en"])
image = example["image"]

For a repeatable experiment, supply the exact Hub commit as revision.

hf download haxiaofeng/CityAgentBench-tasks --repo-type dataset \
  --local-dir downloads/CityAgentBench-tasks

Evaluation use

There are no public answer labels in this preview, so it does not support computing an official accuracy score. A complete evaluation release must bind the code, runtime, scene, reviewed tasks, budgets and scoring. Future development sets may provide explicit public labels; hidden-test answers will be managed separately from agent-visible material.

The current GitHub repository is being expanded from its website source into the collaboration/evaluation entry point. Runtime packaging and formal test activation are ongoing.

Source and limitations

These are rendered simulated-city observations, not photographs establishing current real-world business names, hours or accessibility. Native layouts and task wording can require revision. The 35 items are one historical content batch, not a representative sample of the whole city. No training or test split membership beyond this historical preview is asserted.

No project-wide dataset license has yet been selected by the owners. Public availability does not provide a blanket reuse grant. Source assets retain their applicable terms. Publication date: 2026-10-23.

中文说明

首批提供 35 道历史 L1 题与 35 张原图，便于查看任务形式和开发数据读取流程。 本批保留 2026-09-28 的单图、零动作规则；当前 L1 已允许移动，不能混用成绩。 这些资料不代表正式测试集、当前 saved193 场景验收或新一轮人工审核通过。 完整评测环境、冻结规则和可执行题目将按版本另行发布。三维场景资源放在独立 Dataset 仓库。