DORA is released for non-commercial research; each part keeps the license of its source (see LICENSE.md). Access is granted automatically. Your name, affiliation and email are shared with the DORA authors and used only to contact you about the dataset (e.g. corrections and new versions).

DORA: Can LLM Agents Respond to Disasters?

Data release (v1.1) of DORA (Disaster Operational Response Agent benchmark), NeurIPS 2026 Evaluations & Datasets Track. Paper: arXiv:2605.11633. Code, tools and evaluator: github.com/Junjue-Wang/DORA.

DORA has 515 expert-authored tasks over 45 real-world disaster events, each with a heterogeneous data manifest (optical, SAR and multispectral imagery, DEM/DSM, thermal and social vector layers), an expert tool plan, an executed gold trajectory over a 108-tool MCP library, and a structured final answer.

tasks/          5 JSON files, one per dimension (T1 perception & assessment, T2 spatial relations,
                T3 operational planning, T4 temporal evolution, T5 report synthesis), and
                gold_polygons/ with the gold polygons of the 22 polygon-path answers
images/         input layers referenced by the tasks (paths in tasks are relative to images/)
checkpoints/    13 perception models (*.safetensors) used by the seg.* tools, with their licenses
manifest.json   size and SHA-256 of every file
LICENSE.md      licenses and required attributions, per source

Versions

v1.1 (this revision, main) updates the task specifications, gold answers and tools; use it with the current code. Gold answers are computed from the human labels. Scores are not directly comparable with v1.0.
v1.0 (revision v1.0) is the release that accompanies the paper; use it with the code release of the same name (git checkout v1.0).

Usage

Request access with the form on this page (granted automatically), then log in once on the machine that downloads the data with a Read token (a fine-grained token needs "Read access to contents of all public gated repos you can access"):

git clone https://github.com/Junjue-Wang/DORA && cd DORA
hf auth login                           # or set HF_TOKEN
python scripts/prepare_data.py          # downloads this dataset into ./data and verifies it

Licenses

DORA's own annotations (tasks/) are released under CC BY-NC-SA 4.0. Every other file keeps the license of its source, listed in LICENSE.md:

source imagery and layers: CC BY-NC-SA 4.0, CC BY-NC 4.0, CC BY 4.0, CC0 or public domain, depending on the source (xBD, DisasterM3, BRIGHT, NAIP, Maxar Open Data Program, Landslide4Sense, RescueNet, CRASAR-U-DROIDs and others, see LICENSE.md);
PlanetScope crops (images/planet/data/houston*/houston*.tif): CC BY-SA 4.0 — Planet Disaster Data (Hurricane Harvey 2017), © 2017 Planet Labs Inc.;
checkpoints: DINOv3-based models under the DINOv3 License (Built with DINOv3), the SegFormer models under the NVIDIA Source Code License, the SpaceNet 7 HRNet under Apache-2.0 and the GVLM landslide model under CC BY-NC-SA 4.0 (details and training-data terms in LICENSE.md).

Use is limited to non-commercial purposes by several of these licenses. The GVLM-CD imagery and ten downsampled RescueNet images used in the paper are not redistributed; scripts/prepare_data.py downloads them from their official sources (see LICENSE.md).

Intended and unacceptable uses

DORA is intended for humanitarian and disaster-response research and for evaluating AI agents. The following uses are not acceptable:

military or intelligence targeting;
surveillance of individuals or communities;
discriminatory pricing, underwriting or redlining.

The release contains scenes of historical events only and no live data-acquisition interface. Population is included only as aggregate rasters; there are no personal records, parcel-level attributes or policyholder data.

Reporting misuse

Report suspected misuse to kingdrone@edu.k.u-tokyo.ac.jp. Reported uses are reviewed against this policy, and access to the dataset can be revoked.

Citation

@inproceedings{wang2026dora,
  title={Can LLM Agents Respond to Disasters? Benchmarking Heterogeneous Geospatial Reasoning in Emergency Operations},
  author={Wang, Junjue and Xuan, Weihao and Qi, Heli and Dai, Pengyu and Liu, Kunyi and Chen, Hongruixuan and Zheng, Zhuo and Xia, Junshi and Ermon, Stefano and Yokoya, Naoto},
  booktitle={Proceedings of the Neural Information Processing Systems},
  year={2026}
}

Please also cite the datasets DORA builds on (listed in LICENSE.md).