zhangdw/GPSBench-10pct

🛰️ GPSBench 10% Stratified Sample

A deterministic coordinate-reasoning subset for lightweight GPS and geographic QA experiments

    

GPSBench 10% is a reproducible subset of joey234/GPSBench for quick experiments on GPS coordinate manipulation and coordinate-grounded geographic reasoning.

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

---

This is a 10% stratified sample of the public GPSBench release, not the full benchmark. It keeps the original data/ tree, task files, split files, and JSON schemas so existing GPSBench-style loaders can point at the same paths after download.

✨ Why This Subset?

GPSBench evaluates whether language models can reason over GPS coordinates directly, covering both pure coordinate operations and applied geographic knowledge. The full release is useful for final evaluation; this compact subset is intended for faster prompt iteration, smoke tests, parser checks, and low-cost model comparisons.

It is designed for questions like:

• Can a model compute distances, bearings, bounding boxes, and coordinate transformations from text alone?
• How does performance differ between pure GPS operations and applied geographic reasoning?
• Can an evaluation pipeline be debugged against the original GPSBench folder layout before a full run?
• Can every sampled file be regenerated with a fixed seed and documented stratification policy?

📦 Dataset at a Glance

6,201 sampled split examples	18 task entries	2 tracks	91 source-matched data files
3,756 train examples	609 dev examples	1,836 test examples	10% sampling fraction

Track coverage

Track	Train	Dev	Test	Total	Task entries
Applied	1,920	303	918	3,141	9
Pure GPS	1,836	306	918	3,060	9
Total	3,756	609	1,836	6,201	18

Task inventory

Track	Tasks
Pure GPS	Format Conversion · Coordinate System Transformation · Distance Calculation · Bearing Computation · Coordinate Interpolation · Area & Perimeter · Bounding Box · Route Geometry · Relative Position
Applied	Place Association · Name Disambiguation · Relative Position · Proximity · Route Analysis · Boundary Analysis · Spatial Patterns · Missing Data Inference · Terrain Classification

🚀 Quick Start

Load one GPSBench split file directly from the Hub:

from datasets import load_dataset

base = "https://huggingface.co/datasets/zhangdw/GPSBench-10pct/resolve/main"
ds = load_dataset(
    "json",
    data_files=f"{base}/data/track_pure_gps/splits/task4_distance_calculation_test.json",
    split="train",
)

print(len(ds))
print(ds[0]["question"])
print(ds[0]["ground_truth"])

Download the full source-compatible tree:

hf download zhangdw/GPSBench-10pct \
  --type dataset \
  --local-dir data/gpsbench-10pct

After download, GPSBench-style paths are preserved:

data/track_pure_gps/splits/task4_distance_calculation_test.json
data/track_applied/splits/task1_place_association_train.json
data/splits_summary.json

🗂️ Files

Path	Description
data/splits_summary.json	Source-style split summary with sampled counts.
data/track_pure_gps/*.json	10% samples of the pure GPS task-level JSON files.
data/track_pure_gps/splits/*.json	10% samples of pure GPS train/dev/test split files plus split metadata.
data/track_applied/*.json	10% samples of the applied task-level JSON files.
data/track_applied/splits/*.json	10% samples of applied train/dev/test split files plus split metadata.
metadata/counts.json	Sampling seed, source revision, track counts, split counts, and per-task counts.
scripts/build_gpsbench_10pct_dataset.py	Exact script used to regenerate this subset from joey234/GPSBench.

🧭 Example Workflows

Read the sampled split summary

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="zhangdw/GPSBench-10pct",
    repo_type="dataset",
    filename="data/splits_summary.json",
)

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

print(len(summary["tasks"]))
print(sum(task["total_examples"] for task in summary["tasks"]))

Inspect all test files for one track

from pathlib import Path
import json

root = Path("data/gpsbench-10pct/data/track_applied/splits")
for path in sorted(root.glob("*_test.json")):
    rows = json.loads(path.read_text(encoding="utf-8"))
    print(path.name, len(rows), rows[0]["task"])

Regenerate the subset

python scripts/build_gpsbench_10pct_dataset.py \
  --source source/joey234-GPSBench \
  --output gpsbench-10pct \
  --fraction 0.10 \
  --seed 20260606

🧪 Sampling Policy

Setting	Value
Source dataset	joey234/GPSBench
Source revision	64dfdd9a96d9813c4449bfe4e395ff8524b69a3e
Sampling fraction	0.10
Sampling seed	20260606
Preserved structure	Original data/ tree, task files, split files, and split metadata.
Stratification	Per JSON file, using available task, subtask, difficulty, track, task-id, and selected ground_truth category fields.

The stratifier uses available fields including task, subtask, difficulty, metadata.track, metadata.track_task_id, metadata.original_task, and selected ground_truth fields such as type, region, granularity_level, noise_type, difficulty, terrain_type, and relation.

✅ Integrity Notes

• The data/ file tree is intentionally identical to the source dataset: 91 data files in both releases.
• Every array-style JSON data file is sampled at exactly round(0.10 * source_file_rows).
• The split-file total is 6,201 examples. Task-level JSON files are also sampled to 10%, but they are not added to the split total to avoid double-counting alternative source surfaces.
• The source GPSBench release has task-level JSON files and split JSON files that are not always the same instance set. This subset therefore samples each JSON data file independently while preserving schema and path compatibility.

📚 Citation

Please cite the original GPSBench paper when using this subset:

@article{gpsbench2025,
  title   = {GPSBench: Do Large Language Models Understand GPS Coordinates?},
  author  = {Truong, Thinh Hung and Lau, Jey Han and Qi, Jianzhong},
  journal = {arXiv preprint arXiv:2602.16105},
  year    = {2025}
}