Earth-Bench-OW

Earth-Bench-OW evaluates Earth observation agents on 248 open-world tasks. Agents acquire or discover the required observations at runtime, execute scientific workflows, and produce open-ended answers supported by execution evidence.

This repository releases the Earth-Bench-OW regime described in Earth-Agent-Pro, building on the task cores introduced by Earth-Agent. The Pro paper organizes 248 matched task cores into 744 questions across IF, AP, and OW. This repository contains the 248 OW records. The companion prepared-input release is Earth-Bench-V2.

Papers and code

• Earth-Agent: Unlocking the Full Landscape of Earth Observation with Agents (ICLR 2026).
• Earth-Agent-Pro: Towards Real-World Full-Chain Earth Observation with Agents.
• Earth-Agent code and evaluation framework.

Task coverage

Modality	Task IDs	Task cores
Spectrum	1–100	100
Products	101–188	88
RGB	189–248	60
Total	1–248	248

Task IDs match the companion release. The number of local files reflects this snapshot; it is not the number of questions or evaluation instances.

Repository layout

Earth-Bench-OW/
├── README.md
├── LICENSE
├── THIRD_PARTY_NOTICES.md
├── task_index.jsonl
├── question_open.json
├── data/question189/ ... data/question248/
├── predownload_data/
├── model_results.csv
├── model_out/
└── tools/

Component	Contents
question_open.json	A JSON object keyed by task IDs "1" through "248"; 248 records.
dialogs	User request, reference assistant/tool interactions, and an open-ended final answer.
data (annotation field)	Acquisition specifications for Spectrum/Products tasks, or a local input directory for RGB tasks.
Acquisition specifications	Region bounds, data collection IDs, selected bands, requested dates, resolution, and optional sampling/scaling settings.
evaluation	One entry per task with its reference answer, metric, and AP-style regime label.
data/ (directory)	254 local RGB input files for tasks 189–248.
predownload_data/	4,223 cached source files: 3,167 GeoTIFFs and 1,056 HDF files; approximately 92.7 GB.
tools, files	Reserved fields; both are null in this release.

model_results.csv is a semicolon-delimited expert-model response table with columns file_path, model, text_prompt, bbox, and result. model_out/ holds 26 stored output masks used by the perception tools. These files support replay and are not leaderboard scores.

task_index.jsonl provides one derived row per task for browsing: question_id, modality, question, reference_answer, and choices. The original JSON file remains the source of truth for trajectories, acquisition specifications, and evaluation details. This release provides a single test index and no official training split.

Download

Install huggingface_hub, then download from the Hub. The full snapshot occupies approximately 92.9 GB.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="LvZTTTT/Earth-Bench-OW",
    repo_type="dataset",
    local_dir="benchmark",
)

For annotations only:

snapshot_download(
    repo_id="LvZTTTT/Earth-Bench-OW",
    repo_type="dataset",
    local_dir="benchmark",
    allow_patterns=["README.md", "question_open.json", "task_index.jsonl"],
)

Read the annotations

import json
from pathlib import Path

questions = json.loads(Path("benchmark/question_open.json").read_text())
example = questions["1"]
request = example["dialogs"][0]["content"]
reference_trajectory = example["dialogs"][1:-1]
reference_answer = example["dialogs"][-1]["content"]
evaluation = example["evaluation"]

Launch tools from a working directory where benchmark/ resolves to the downloaded repository. The bundled MCP configuration uses tools/... paths; either install those scripts under your framework's tools/ directory or update the configuration to their actual paths. Configure the output directory consistently with the reference benchmark/out/... paths.

Evaluation and reproducibility

Use the papers for the regime definitions, final-answer evaluation, and trajectory metrics. The original Earth-Agent framework provides evaluation code; integration of this snapshot requires selecting the correct annotation file, regime, data paths, and tool implementation. This dataset repository does not include an end-to-end evaluation runner.

• OW has no candidate-answer field. Reference final answers can be strings or numbers; preserve their meaning when evaluating.
• Runtime paths such as benchmark/data/<region>_<band>_<date>.tif in the trajectories represent acquired or processed files. They need not exist before execution.
• predownload_data/ contains cached source observations. For a cached replay, explicitly connect your acquisition tools to this directory. For live OW evaluation, perform runtime acquisition/discovery and report the data source and access date.
• Keep acquisition specifications, reference trajectories, final answers, model_results.csv, and model_out/ in the evaluator's trusted environment. Expose only the inputs and tool responses prescribed by the selected evaluation protocol.
• The bundled Data MCP server (tools/Data.py) implements download_product and apply_scale_offset. Set EARTH_ENGINE_PROJECT to your own project and configure Earth Engine authentication. Reference trajectories also call download_mod021km, download_myd021km, and download_aster_l1t; these three helpers require a compatible external implementation. The cache alone does not implement those calls.
• The included perception tools replay stored expert-model responses. Report this mode explicitly; live expert-model inference requires a separate implementation and model weights.

Intended use

Use this benchmark to evaluate scientific Earth observation workflows, tool selection, argument grounding, and final-answer quality. The released trajectories and answers are evaluation references. Disclose any use of them for training or prompt construction before reporting benchmark performance.

License and sources

Benchmark annotations and release documentation are licensed under CC BY-NC 4.0. Attribute the benchmark and cite the relevant papers. Third-party satellite products, RGB imagery, and inherited code retain their original terms; see THIRD_PARTY_NOTICES.md. The benchmark license does not replace those terms.

Citation

Please cite both papers when using these matched benchmark releases.

@article{feng2025earthagent,
  title   = {Earth-Agent: Unlocking the Full Landscape of Earth Observation with Agents},
  author  = {Feng, Peilin and Lv, Zhutao and Ye, Junyan and Wang, Xiaolei and Huo, Xinjie and Yu, Jinhua and Xu, Wanghan and Zhang, Wenlong and Bai, Lei and He, Conghui and Li, Weijia},
  journal = {arXiv preprint arXiv:2509.23141},
  year    = {2025},
  url     = {https://arxiv.org/abs/2509.23141}
}

@article{lv2026earthagentpro,
  title   = {Earth-Agent-Pro: Towards Real-World Full-Chain Earth Observation with Agents},
  author  = {Lv, Zhutao and Dang, Chenhao and Feng, Yi and Gong, Yanpei and Wang, Xiaolei and Ye, Junyan and He, Conghui and Li, Weijia},
  journal = {arXiv preprint arXiv:2609.12533},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.12533}
}