🗺️ GABench Data

Hub-hosted GIS artifacts for GABench / GeoAgentBench evaluation

    

GABench Data stores the large benchmark table, GIS input layers, and reference outputs used by the GABench / GeoAgentBench evaluation code.

Quick Start · At a Glance · Files · Benchmark Table · Citation

---

This dataset is an artifact mirror, not a standalone benchmark implementation. The Hugging Face repository stores the files that were previously tracked with Git LFS; the benchmark code and evaluation logic live in the companion GitHub repository.

✨ Why This Dataset?

GABench / GeoAgentBench evaluates agents on realistic GIS analysis tasks that combine natural-language instructions, geospatial data layers, toolchain planning, and generated map or table outputs. The data artifacts are too large and heterogeneous for a lightweight code checkout, so this Hub repository keeps the file tree separately while preserving the relative paths expected by the evaluation code.

It is designed for questions like:

• How do I restore the GABench benchmark/ and dataset/ folders without pulling large Git LFS blobs through Git?
• Which GIS domains, input layers, toolchain lengths, and reference outputs are covered by the benchmark table?
• Can I inspect the task descriptions and expected output files before running the evaluator?
• Can the code repository stay small while the benchmark artifacts remain versioned and downloadable?

📦 Dataset at a Glance

53 benchmark tasks	6 GIS domains	190 artifact files	2.60 GB Hub storage
188 dataset/ files	50 reference outputs	3-17 toolchain steps	1 benchmark table

Domain coverage

Domain	Tasks
Raster Spatial Analysis	16
Geostatistical Analysis	13
Hydrological Analysis	11
Vector Spatial Analysis	7
3D Modeling and Analysis	4
Spatial Data Management	2
Total	53

File-format coverage

File family	Formats	Count
Raster and gridded data	.tif, .nc	34
Vector GIS data	.geojson, .shp, .dbf, .shx, .prj, .gpkg	95
Tables and spreadsheets	.csv, .xlsx	9
Reference visual outputs	.png	49
Graph or serialized objects	.graphml, .pkl	2
Manifest	.txt	1

🚀 Quick Start

Download the benchmark artifacts into a local checkout of the code repository:

uvx --from huggingface_hub hf download zhangdw/GABench \
  --repo-type dataset \
  --local-dir . \
  --include 'benchmark/**' \
  --include 'dataset/**' \
  --include 'metadata/**'

After download, the original relative paths expected by the benchmark code are restored:

benchmark/benchmark.csv
dataset/<GIS input files>
dataset/result/<reference outputs>
metadata/lfs-files.txt

Read the benchmark table directly from the Hub:

import pandas as pd

url = "https://huggingface.co/datasets/zhangdw/GABench/resolve/main/benchmark/benchmark.csv"
df = pd.read_csv(url, encoding="utf-8-sig")

print(len(df))
print(df[["ID", "Domain", "Task Description", "Toolchain Length", "Result"]].head())

Download only the reference outputs:

uvx --from huggingface_hub hf download zhangdw/GABench \
  --repo-type dataset \
  --local-dir gabench-results \
  --include 'dataset/result/**'

🗂️ Files

Path	Description
benchmark/benchmark.csv	Canonical benchmark task table with 53 GIS analysis tasks.
dataset/	GIS input layers plus reference outputs used by the benchmark.
dataset/result/	Expected result artifacts, mostly .png maps plus one .csv output.
metadata/lfs-files.txt	Manifest of files migrated from the Git LFS-backed source tree.

🧱 Benchmark Table

benchmark/benchmark.csv is the primary task index. The logical columns are:

Column	Description
ID	Stable task identifier.
Domain	GIS task family, such as raster, vector, hydrological, or geostatistical analysis.
Task Description	Natural-language instruction for the agent or evaluator.
Data Description	Input files and field-level hints needed by the task.
Drawing Style	Visualization requirements for map-style outputs.
Toolchain Length	Number of expected toolchain steps.
Toolchain JSON	Structured representation of the target GIS workflow.
Result	Expected output filename under dataset/result/ or related output path.
Layers	Referenced GIS layers.

The source CSV keeps several trailing empty columns. They are preserved here to avoid changing the upstream-compatible table layout.

🧭 Example Workflows

Inspect task distribution by domain

import pandas as pd

df = pd.read_csv("benchmark/benchmark.csv", encoding="utf-8-sig")
print(df.groupby("Domain").size().sort_values(ascending=False))

Find tasks that require longer toolchains

long_tasks = df[df["Toolchain Length"].astype(int) >= 10]
print(long_tasks[["ID", "Domain", "Toolchain Length", "Result"]])

List referenced result files

expected = sorted(df["Result"].dropna().unique())
for name in expected[:20]:
    print(name)

🔗 Use With Code

This dataset is intended to be paired with the code repository rather than loaded as a conventional train/test split.

Component	Location
Upstream project	GeoX-Lab/GABench
Companion code fork	zhangdw156/GABench
Code branch	feat/leo-260609
Data artifact repository	zhangdw/GABench on Hugging Face

The GitHub repository keeps the benchmark and evaluation code; this Hugging Face dataset stores the benchmark table, GIS inputs, and reference outputs.

✅ Intended Use

This repository is useful for:

• restoring the GABench / GeoAgentBench data tree for evaluation runs;
• inspecting benchmark tasks, domains, required layers, toolchains, and outputs;
• separating large GIS artifacts from a lightweight code checkout;
• reproducing the artifact state used by the feat/leo-260609 companion branch.

It is not meant to replace the official benchmark code, define a new evaluation protocol, or serve as a general-purpose geospatial training corpus.

⚖️ License

The upstream repository did not expose a clear license file at the time this dataset card was created, so the Hub metadata is marked as other rather than inferred. Before redistributing, modifying, or using these files in downstream releases, check the original GABench / GeoAgentBench project and the terms of any source geospatial datasets referenced by individual tasks.

🛠️ Maintenance Notes

• The path layout intentionally mirrors the source tree expected by the benchmark code.
• metadata/lfs-files.txt records the files that were migrated out of Git LFS.
• README-only edits should not imply a new data version unless the benchmark table, input artifacts, or reference outputs change.
• If upstream GABench / GeoAgentBench changes its task table or GIS files, update the manifest and this card together.

📚 Citation

If you use this artifact mirror, cite the original GeoAgentBench paper for benchmark design and evaluation methodology, and cite this Hub dataset for data provenance.

@misc{yu2026geoagentbenchdynamicexecutionbenchmark,
  title         = {GeoAgentBench: A Dynamic Execution Benchmark for Tool-Augmented Agents in Spatial Analysis},
  author        = {Bo Yu and Cheng Yang and Dongyang Hou and Chengfu Liu and Jiayao Liu and Chi Wang and Zhiming Zhang and Haifeng Li and Wentao Yang},
  year          = {2026},
  eprint        = {2604.13888},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2604.13888}
}

You may also cite this Hugging Face artifact mirror when data provenance matters:

@misc{gabench_data2026,
  title        = {GABench Data},
  author       = {Dawei Zhang},
  year         = {2026},
  howpublished = {Hugging Face Dataset},
  url          = {https://huggingface.co/datasets/zhangdw/GABench}
}

---

A path-compatible Hugging Face artifact mirror for GIS-agent benchmark evaluation.