zhangdw/USTBench-ST-Planning-10pct

🏙️ USTBench ST Understanding + Planning 10% Sample

A deterministic, stratified QA subset for urban spatial-temporal reasoning and planning

    

USTBench ST Planning 10% is a public, deterministic subset of Haruto2099/USTBench-Dataset, focused on process-based spatial-temporal understanding and planning QA.

Quick Start · At a Glance · Files · Sampling · Citation

---

This is a sampled subset, not the full USTBench release. It keeps only st_understanding and planning, uses a deterministic 10% sampling policy, and preserves provenance fields so each row can be traced back to the original USTBench source file and index.

✨ Why This Subset?

USTBench is useful for evaluating urban spatial-temporal reasoning, but the full QA tree can be heavy for quick experiments, smoke tests, and prompt iteration. This dataset keeps a compact, stratified slice of the two process-based QA subsets most directly aligned with spatial-temporal understanding and planning.

It is designed for questions like:

• Can a model answer urban spatial-temporal reasoning questions from a lightweight sample?
• How do results differ between st_understanding and planning tasks?
• Can experiments run against both Hugging Face JSONL and USTBench-compatible file layouts?
• Can every sampled row be traced back to its original task, file, and index?

📦 Dataset at a Glance

4,131 sampled QA cases	2 kept subsets	10% sampling fraction	20260606 sampling seed
2,631 st_understanding	1,500 planning	9 source task folders	2 file layouts

Included subsets

Subset	Sampled cases	Stratification
st_understanding	2,631	Source task × spatial/temporal type × relation.
planning	1,500	Source task.
Total	4,131	Deterministic 10% sample.

Task coverage

congestion_prediction · next_poi_prediction · poi_placement · road_planning · route_planning · socio_ecomic_prediction · traffic_od_prediction · traffic_signal_control · urban_planning

The original release uses the folder name socio_ecomic_prediction. This sample also includes a compatibility alias socio_economic_prediction for the same sampled rows because USTBench example scripts use that spelling. The alias is not counted in metadata/counts.json totals.

🚀 Quick Start

Load the combined JSONL file:

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="https://huggingface.co/datasets/zhangdw/USTBench-ST-Planning-10pct/resolve/main/data/all.jsonl",
    split="train",
)

print(len(ds))
print(ds[0].keys())

Load the two subsets separately:

from datasets import load_dataset

base = "https://huggingface.co/datasets/zhangdw/USTBench-ST-Planning-10pct/resolve/main"
files = {
    "st_understanding": f"{base}/data/st_understanding.jsonl",
    "planning": f"{base}/data/planning.jsonl",
}

data = load_dataset("json", data_files=files)
print(data["st_understanding"].num_rows)
print(data["planning"].num_rows)

Use the original USTBench-style tree after download:

hf download zhangdw/USTBench-ST-Planning-10pct \
  --type dataset \
  --local-dir data/ustbench-st-planning-10pct

question_answering/Data/<task>/st_understanding_QA.json
question_answering/Data/<task>/planning_QA.json

🗂️ Files

Path	Description
data/st_understanding.jsonl	Hugging Face-friendly JSONL for the sampled ST-understanding cases.
data/planning.jsonl	Hugging Face-friendly JSONL for the sampled planning cases.
data/all.jsonl	Concatenation of both sampled subsets.
metadata/counts.json	Source counts, sample counts, sampling seed, and stratum-level details.
question_answering/Data/<task>/*_QA.json	USTBench-compatible tree for scripts expecting the original layout.

🧭 Example Workflows

Inspect provenance fields

row = ds[0]
provenance = {k: v for k, v in row.items() if k.startswith("_ustbench_")}
print(provenance)

Each sampled row includes _ustbench_ fields for source file, source task, subset, original index, sampling cell, seed, and sampling fraction.

Count rows by subset and task

import pandas as pd

df = ds.to_pandas()
print(df.groupby(["_ustbench_subset", "_ustbench_source_task"]).size())

Read the deterministic sampling manifest

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="zhangdw/USTBench-ST-Planning-10pct",
    repo_type="dataset",
    filename="metadata/counts.json",
)

with open(path, "r", encoding="utf-8") as f:
    counts = json.load(f)

print(counts["sampling_seed"])
print(counts["total_sampled_cases"])

🧪 Sampling Policy

Setting	Value
Source dataset	Haruto2099/USTBench-Dataset
Sampling fraction	0.1
Sampling seed	20260606
Kept subsets	st_understanding, planning
st_understanding strata	Source task, spatial/temporal type, spatial/temporal relation.
planning strata	Source task.

The goal is stable, reproducible coverage rather than a new benchmark protocol. Use the full USTBench release when full-scale reporting is required.

✅ Intended Use

This subset is useful for:

• quick prompt and evaluator iteration before running full USTBench;
• lightweight spatial-temporal QA experiments;
• comparing process-based st_understanding and planning behavior;
• reproducing experiments with explicit sampling provenance;
• testing code that expects either JSONL or the original USTBench directory layout.

⚖️ License

This sampled dataset follows the CC-BY-4.0 license metadata of the upstream source dataset. Check the original USTBench release for source-specific terms and citation expectations.

📚 Citation

If you use this sample, cite the original USTBench paper for the benchmark design and cite this deterministic subset when the 10% sampling procedure matters for reproducibility:

@misc{lai2025ustbenchbenchmarkingdissectingspatiotemporal,
  title         = {USTBench: Benchmarking and Dissecting Spatiotemporal Reasoning of LLMs as Urban Agents},
  author        = {Siqi Lai and Yansong Ning and Zirui Yuan and Zhixi Chen and Hao Liu},
  year          = {2025},
  eprint        = {2505.17572},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2505.17572}
}

You may also cite this deterministic subset when the sampling policy matters for reproducibility:

@misc{ustbench_st_planning_10pct2026,
  title        = {USTBench ST Understanding + Planning 10% Sample},
  author       = {Dawei Zhang},
  year         = {2026},
  howpublished = {Hugging Face Dataset},
  url          = {https://huggingface.co/datasets/zhangdw/USTBench-ST-Planning-10pct}
}

---

A compact, provenance-preserving slice of USTBench for urban spatial-temporal QA experiments.