PolarTools

A benchmark for scientific analysis of polar Earth observations.

Tool repository · Question index · Dataset manifest · Task categories

PolarTools contains 717 numbered questions, including 710 active tasks and 7 retired historical tasks. Tasks cover polar remote sensing, scientific preprocessing, spatial measurement, temporal change, and geospatial mapping, using SAR and optical imagery, UAV observations, masks, vector data, NetCDF products, and professional references.

Dataset overview

Property	Contents
Questions	Question1/ through Question717/
Active tasks	710
Retired tasks	Question390, Question400, Question405, Question411, Question417, Question423, Question429
Language	English
Inputs	SAR and optical observations, UAV orthomosaics, masks, vector regions, NetCDF products, sensor metadata, scientific processing packages, and reference documents
Outputs	Values, structured JSON, masks, geospatial rasters, GeoJSON, charts, and maps

Task categories

Primary category	Active tasks	Capabilities
Recognition, localization, and knowledge interpretation	32	Target localization and evidence retrieval
Measurement and counting	125	Target counting, spatial measurement, and observation statistics
Segmentation and spatial structure analysis	145	Target delineation, mask topology, and spatial structure
Change analysis	256	Temporal differences, spatial trajectories, and observation changes
Quantitative remote sensing analysis	29	Product analysis, sensor units, and quantitative diagnostics
Preprocessing and cartography	123	Input preparation, grid alignment, sensor processing, and georeferenced mapping
Total	710

Segmentation and spatial structure analysis comprises 61 raw-image segmentation tasks and 84 supplied-mask structure and statistics tasks. Recognition, localization, and knowledge interpretation includes 10 evidence-retrieval tasks.

Dataset composition

Folder range	Task family
Question1–Question600	Original polar task pool
Question601–Question610	Evidence retrieval from professional references
Question611–Question613	Cartography
Question614–Question643	Additional cartography
Question644–Question673	Remote-sensing preprocessing
Question674–Question693	Raw-image iceberg segmentation
Question694–Question717	Multi-observation iceberg identity, tracking, and cartography

Data structure

PolarTools/
├── README.md
├── tasks.jsonl
├── task_index.md
├── taxonomy.json
├── manifest.json
├── Question1/
│   ├── task.json
│   ├── README.md
│   ├── 01_...tif
│   └── 02_...geojson
├── Question2/
│   └── ...
└── Question717/
    ├── task.json
    ├── README.md
    ├── input_manifest.json
    └── Question76/

Download

Complete dataset

python -m pip install huggingface_hub
hf download PolarTools/PolarTools \
  --repo-type dataset \
  --local-dir ./PolarTools-Benchmark

Individual question

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="PolarTools/PolarTools",
    repo_type="dataset",
    local_dir="./PolarTools-Benchmark",
    allow_patterns=[
        "README.md", "tasks.jsonl", "manifest.json", "taxonomy.json",
        "Question1/**",
    ],
)

Task metadata

from datasets import load_dataset

archive = load_dataset("PolarTools/PolarTools", "metadata", split="archive")
active = archive.filter(lambda row: row["active"])
print(len(archive), len(active))

Use a question

import json
from pathlib import Path

question_dir = Path("PolarTools-Benchmark/Question1")
task = json.loads((question_dir / "task.json").read_text())
input_files = [question_dir / name for name in task["input_files"]]

assert all(path.is_file() for path in input_files)
print(task["id"], task["title"])
print(task["question"])

Evaluation settings

• General Solving (GS): general reasoning with optional self-written code.
• Tool-Augmented (TA): general reasoning with specialist tools.