K-MetBench: A Multi-Dimensional Benchmark for Fine-Grained Evaluation of Expert Reasoning, Locality, and Multimodality in Meteorology

K-MetBench is a multi-dimensional benchmark for evaluating meteorology models across accuracy, reasoning quality, geo-cultural alignment, and fine-grained domain coverage.

The public eval protocol uses only the explicit advanced benchmark and the explicit reasoning benchmark followed by LLM-as-a-judge evaluation. The implicit split may be distributed with the dataset, but it is not part of the public eval kit.

Dataset Summary

• Total Questions: 1774
• Total Image References: 151 (59 question images, 92 choice images)
• Modality Split: text-only 1692, multimodal 82
• Reasoning Subset: 141
• Geo-Cultural Subset: 73
• Parts: Part 1: 373, Part 2: 332, Part 3: 359, Part 4: 376, Part 5: 334
• Format: JSON file with relative image paths under data/images/

Data Format

Each sample contains:

Field	Type	Description
id	int	Stable item identifier
question.text	string	Question text
question.image	string	Relative path to a question image, if present
choices[].text	string	Choice text
choices[].image	string	Relative path to a choice image, if present
answer	int	Zero-based correct choice index
source	string	Exam session source tag
source_id	int	Original source-local item id
rationale	string	Expert-verified reasoning text when available
korean	bool	Geo-cultural subset flag
multimodal	bool	Multimodal subset flag
part	int	Official part number (1-5)
category	object	Subject/topic metadata

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files="https://huggingface.co/datasets/soyeonbot/K-MetBench/resolve/main/data/kmetbench.json",
    split="test",
)

sample = dataset[0]
print(sample["question"]["text"])
print(sample["answer"])

Viewing Referenced Images

import requests
from io import BytesIO
from PIL import Image

image_rel_path = sample["question"]["image"]
image_url = "https://huggingface.co/datasets/soyeonbot/K-MetBench/resolve/main/data/images/" + image_rel_path
image = Image.open(BytesIO(requests.get(image_url, timeout=30).content))
image.show()

Running the Public Eval Kit

pip install -r requirements-eval.txt
python scripts/eval.py run --list-model-configs
python scripts/eval.py run --model-config <model_config> --prompt-type advanced --explicit-data-file data/kmetbench.json --image-root data/images
python scripts/eval.py run --model-config <model_config> --prompt-type reasoning --explicit-data-file data/kmetbench.json --image-root data/images
python scripts/eval.py judge --model <model> --predictions <explicit_reasoning_json> --explicit-data-file data/kmetbench.json

License

This dataset is released under CC BY-NC-SA 4.0.

Contact

For questions about the dataset, contact Soyeon Kim (soyeon.k@kaist.ac.kr).

Citation

@inproceedings{kim2026kmetbench,
  title = {K-MetBench: A Multi-Dimensional Benchmark for Fine-Grained Evaluation of Expert Reasoning, Locality, and Multimodality in Meteorology},
  author = {Kim, Soyeon and Kang, Cheongwoong and Lee, Myeongjin and Chang, Eun-Chul and Lee, Jaedeok and Choi, Jaesik},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  url = {http://arxiv.org/abs/2604.24645}
}