GMRID v3: Global Maritime and Supply-Chain Risk Intelligence Dataset

All authorship and attribution belong to the original creators. This is a mirror of the dataset from inflaton/llms-at-edge hosted on Hugging Face for accessibility. The original repository does not specify a license; please contact the authors for licensing terms before commercial use.

Overview

GMRID v3 is a supply-chain disruption news classification dataset. Each row is a real-world incident report (headline + details) labeled with one of 8 disruption categories. The dataset was introduced in:

LLMs at the Edge: Performance and Efficiency Evaluation with Ollama on Diverse Hardware IJCNN 2025 (Paper ID: 1443) GitHub: inflaton/llms-at-edge

Task

Single-label classification into 8 categories:

Category	Train	Test
Weather	—	366
Administrative Issue	—	333
Accident	—	191
Worker Strike	—	178
Terrorism	—	60
Human Error	—	9
Others	—	5
Cyber Attack	—	4

The label column is Summarized_label. A finer-grained Category column provides subcategories (e.g., "Flooding" under Weather, "Port Congestion" under Administrative Issue). The mapping is defined in categories.json.

Splits

Split	Rows
Train	4,594
Test	1,147

Columns

Column	Description
id	Unique row identifier
Headline	Short incident headline
Details	Full incident description
Severity	Severity level (Critical, Moderate, etc.)
Region	Geographic region
Datetime	Incident timestamp
lat, lon	Coordinates (when available)
maritime_label	Whether the incident is maritime-related
found_ports	Ports mentioned in the text
contains_port_info	Boolean: port info present
if_labeled	Whether the row was manually labeled
Headline_Details	Concatenated headline + details
Year, Month, Week	Temporal features
Details_cleaned	Preprocessed/cleaned details text
Category	Fine-grained incident category
Summarized_label	Coarse 8-class label (primary target)
gpt-4o_label	GPT-4o predicted label (for reference)

Evaluation Metric

Per the original paper: weighted F1 over the 8-class Summarized_label. Macro-F1, exact-match, and per-class P/R/F1 are also commonly reported.

Citation

If you use this dataset, please cite the original work:

@inproceedings{llms_at_edge_2025,
  title={LLMs at the Edge: Performance and Efficiency Evaluation with Ollama on Diverse Hardware},
  booktitle={International Joint Conference on Neural Networks (IJCNN)},
  year={2025},
  note={Paper ID: 1443},
  url={https://github.com/inflaton/llms-at-edge}
}