BoundaryBench is a benchmark dataset of 13,000 synthetic coordinate queries across 50 U.S. states plus D.C., designed to evaluate geographic boundary resolution systems. Points are stratified by distance-to-boundary difficulty tiers and complex polygon geometries.

Points: 13,000
Layers: county, ZCTA, tract
Difficulties: easy, medium, hard, boundary, edge
Version: v1.0.0
License: CC-BY-4.0
DOI: 10.5281/zenodo.18090588

Independent research; not endorsed by any organization.

How to Load

from datasets import load_dataset

# Load the dataset
ds = load_dataset("nidhipandya/boundarybench")

# Or load parquet files directly
import pandas as pd
df = pd.read_parquet("hf://datasets/nidhipandya/boundarybench/data/boundarybench_v1.parquet")

print(f"Total points: {len(df)}")
print(df.head())

Data Fields

Field	Type	Description
`lat`	float	Latitude (EPSG:4269 / NAD83)
`lon`	float	Longitude (EPSG:4269 / NAD83)
`layer`	string	Administrative layer: `county`, `zcta`, or `tract`
`label_id`	string	Ground-truth region identifier
`difficulty`	string	One of: `easy`, `medium`, `hard`, `boundary`, `edge`
`dist_to_boundary_m`	float	Distance to nearest boundary in meters
`state_fips`	string	State FIPS code
`split`	string	Dataset split: `train`, `dev`, or `test`

Difficulty Stratification

Difficulty	Distance Range	Count
easy	≥500m	4,550
medium	100-500m	3,250
hard	10-100m	3,250
boundary	0-10m	1,300
edge	10-100m	650

Versioning

Version: v1.0.0
Code: GitHub
Zenodo (concept): 10.5281/zenodo.18090588
Zenodo (v1.0.0): 10.5281/zenodo.18090589

Citation

@dataset{pandya2025boundarybench,
  title={BoundaryBench: Evaluating Geographic Boundary Resolution with Calibrated Uncertainty},
  author={Pandya, Nidhi},
  year={2025},
  doi={10.5281/zenodo.18090588},
  url={https://doi.org/10.5281/zenodo.18090588},
  license={CC-BY-4.0}
}

License

CC-BY-4.0

Contact

Author: Nidhi Pandya
Email: nidhipandya1606@gmail.com
GitHub: https://github.com/nidhip1611/boundarybench

Downloads last month: 19