Dataset Card for JamC-QA-V2

English/Japanese

Dataset Summary

JamC-QA-V2 is an expanded version of JamC-QA, a benchmark that evaluates knowledge specific to Japan through multiple-choice questions.

The key expansions in V2 are as follows:

• Added a new category: Japan Self-Defense Forces (jsdf).
• Introduced sub-categories:
  • Currently applied only to the 5 sub-categories under jsdf:
    • Service regulations (regulations)
    • Tri-service common (common)
    • Japan Ground Self-Defense Force (jgsdf)
    • Japan Maritime Self-Defense Force (jmsdf)
    • Japan Air Self-Defense Force (jasdf)

In future updates, sub-categories will be added to other categories as well.

Leaderboard

Evaluation Metric

In our evaluation, the LLM outputs the option string rather than the option label, and the following table shows the proportion of outputs that exactly match the gold option string.

Model	All	culture	custom	regional_identity	geography	history	government	law	healthcare	jsdf
sarashina2-8x70b	0.715	0.714	0.775	0.761	0.654	0.784	0.736	0.632	0.917	0.568
sarashina2-70b	0.712	0.719	0.745	0.736	0.673	0.764	0.764	0.666	0.917	0.523
Llama-3.3-Swallow-70B-v0.4	0.687	0.689	0.775	0.589	0.566	0.776	0.773	0.783	0.854	0.542
RakutenAI-2.0-8x7B	0.626	0.622	0.725	0.617	0.511	0.714	0.709	0.575	0.813	0.523
plamo-100b	0.582	0.602	0.650	0.637	0.504	0.682	0.609	0.515	0.688	0.271
Mixtral-8x7B-v0.1-japanese	0.579	0.602	0.670	0.579	0.493	0.612	0.736	0.545	0.667	0.368
Meta-Llama-3.1-405B	0.564	0.558	0.545	0.484	0.500	0.679	0.646	0.629	0.688	0.465
Nemotron-4-340B-Base	0.558	0.573	0.615	0.511	0.467	0.595	0.727	0.582	0.667	0.426
llm-jp-3.1-8x13b	0.556	0.595	0.635	0.582	0.449	0.589	0.627	0.502	0.625	0.374
Qwen2.5-72B	0.523	0.522	0.595	0.426	0.438	0.606	0.609	0.562	0.688	0.471

The table below shows the evaluation results for JSDF questions broken down by sub-category:

Model	regulations	common	jgsdf	jmsdf	jasdf
sarashina2-8x70b	0.452	0.548	0.548	0.645	0.645
sarashina2-70b	0.484	0.516	0.484	0.581	0.548
Llama-3.3-Swallow-70B-v0.4	0.452	0.581	0.581	0.516	0.581
RakutenAI-2.0-8x7B	0.355	0.645	0.452	0.548	0.613
plamo-100b	0.355	0.290	0.194	0.258	0.258
Mixtral-8x7B-v0.1-japanese	0.387	0.419	0.290	0.387	0.355
Meta-Llama-3.1-405B	0.452	0.613	0.387	0.484	0.387
Nemotron-4-340B-Base	0.387	0.548	0.452	0.419	0.323
llm-jp-3.1-8x13b	0.387	0.419	0.290	0.355	0.419
Qwen2.5-72B	0.419	0.581	0.484	0.419	0.452

Language

Japanese

Dataset Structure

Data Instances

An example from the jsdf category looks as follows:

{
  "qid": "jamcqa-test-jsdf-00004",
  "category": "jsdf",
  "sub_category": "regulations",
  "question": "2026年9月現在、自衛官が敬礼を省略できない場面はどれ?",
  "choice0": "車両、自衛艦、航空機等の操縦に従事しているとき",
  "choice1": "上級者に随従し上級者が敬礼を受けるとき",
  "choice2": "勤務、演習、訓練、作業等に従事し、敬礼がその任務の遂行に支障があるとき",
  "choice3": "怪我などで腕が使えないとき",
  "answer_index": 3,
  "difficulty": 1,
}

Data Fields

• qid (str): A unique identifier for each question.
• category (str): The category of the question.
  • culture, custom, regional_identity, geography, history, government, law, healthcare, and jsdf.
• sub_category (str): The sub-category for fine-grained classification.
  • Currently assigned only to jsdf.
  • Set to an empty string "" if not assigned.
• question (str): The question text.
  • Converted from full-width to half-width characters, excluding katakana characters.
  • Does not contain any line breaks (\n).
  • Leading and trailing whitespace is removed.
• choice{0..3} (str): Four answer options (choice0 to choice3).
  • Converted from full-width to half-width characters, excluding katakana characters.
  • Does not contain any line breaks (\n).
  • Leading and trailing whitespace is removed.
• answer_index (int): The index of the correct answer among choice0 to choice3 (0–3).
• difficulty (int): The number of models that correctly answered the example (0-9).
  • 0: 0 models were able to choose the correct answer.
  • 9: 9 models were able to choose the correct answer.
  • For details, please refer to section 6.2 of [Oka+, 26].

Data Splits

• dev: 4 examples per category, intended for few-shot evaluation. If sub-categories are specified, 4 examples are provided per sub-category.
• test: 2,464 examples in total.

Number of Examples:

Category	dev	test
culture	4	640
custom	4	200
regional_identity	4	397
geography	4	272
history	4	343
government	4	110
law	4	299
healthcare	4	48
jsdf	20	155
total	52	2,464

Licensing Information

• CC-BY-SA-4.0

Usage

Dataset Loading

$ python
>>> import datasets
>>> jamcqa_v2 = datasets.load_dataset('sbintuitions/JamC-QA-V2', 'v2.0')
>>> print(jamcqa_v2)
DatasetDict({
    dev: Dataset({
        features: ['qid', 'category', 'sub_category', 'question', 'choice0', 'choice1', 'choice2', 'choice3', 'answer_index', 'difficulty'],
        num_rows: 52
    })
    test: Dataset({
        features: ['qid', 'category', 'sub_category', 'question', 'choice0', 'choice1', 'choice2', 'choice3', 'answer_index', 'difficulty'],
        num_rows: 2464
    })
})
>>> jamcqa_v2_test = jamcqa_v2['test']
>>> print(jamcqa_v2_test)
Dataset({
    features: ['qid', 'category', 'sub_category', 'question', 'choice0', 'choice1', 'choice2', 'choice3', 'answer_index', 'difficulty'],
    num_rows: 2464
})
>>> print(jamcqa_v2_test[0])
{'qid': 'jamcqa-test-culture-00001', 'category': 'culture', 'sub_category': '', 'question': '「狂った世で気が狂うなら気は確かだ」の名言を残した映画はどれ?', 'choice0': '影武者', 'choice1': '羅生門', 'choice2': '隠し砦の三悪人', 'choice3': '乱', 'answer_index': 3, 'difficulty': 4}
>>> print(jamcqa_v2_test[2312])
{'qid': 'jamcqa-test-jsdf-00004', 'category': 'jsdf', 'sub_category': 'regulations', 'question': '2026年9月現在、自衛官が敬礼を省略できない場面はどれ?', 'choice0': '車両、自衛艦、航空機等の操縦に従事しているとき', 'choice1': '上級者に随従し上級者が敬礼を受けるとき', 'choice2': '勤務、演習、訓練、作業等に従事し、敬礼がその任務の遂行に支障があるとき', 'choice3': '怪我などで腕が使えないとき', 'answer_index': 3, 'difficulty': 1}
>>> 

Evaluation with FlexEval

You can easily use FlexEval (version 18.2 or later, commit 3c996b8 or later) to evaluate the JamC-QA-V2 score by simply replacing commonsense_qa with jamcqa_v2 in the Quickstart guide.

Run Command

flexeval_lm \
  --language_model HuggingFaceLM \
  --language_model.model "sbintuitions/sarashina2.2-0.5b" \
  --language_model.default_gen_kwargs "{ do_sample: false }" \
  --eval_setup "jamcqa_v2" \
  --save_dir "results/jamcqa_v2"

--language_model.default_gen_kwargs "{ do_sample: false }" disables sampling and performs greedy search.

Output

...
2026-10-02 13:56:47.098 | INFO     | flexeval.core.evaluate_generation:evaluate_generation:95 - {'exact_match': 0.23782467532467533, 'finish_reason_ratio-stop': 1.0, 'avg_output_length': 6.985795454545454, 'max_output_length': 93, 'min_output_length': 2}
...

Citation Information

@inproceedings{oka-etal-2026-jamcqa,
  author={Oka, Teruaki and Shibata, Tomohide and Yoshida, Nao},
  title={JamC-QA: A Multiple-Choice Question Answering Benchmark for Japan-Specific Knowledge},
  year={2026},
  month={May},
  booktitle={Proceedings of the Fifteenth Language Resources and Evaluation Conference (LREC 2026)},
  pages={4536--4546},
  doi={10.63317/33xwat5joobd},
}