JamC-QA / README_ja.md
teruo6939's picture
Update README_ja.md
b442b7b verified
|
raw
history blame
5.9 kB

Dataset Card for JamC-QA

英語/日本語

Dataset Summary

このベンチマークは、日本の文化や風習といった日本固有の知識を問う多肢選択式の質問応答ベンチマークです。 文化(culture)、風習(custom)、風土(regional_identity)、地理(geography)、日本史(history)、行政(government)、法律(law)、医療(healthcare)の 8つのカテゴリの問題を含みます。 このベンチマークで高いスコアを達成するには、日本に関する幅広い知識を獲得している必要があります。

Leaderboard

Evaluation Metric

Accuracy

Accuracyは LLM の出力した文字列が正解選択肢の文字列と完全一致した割合です。 本評価では、選択肢に付与されたラベル(0-3)でなく、LLM に選択肢文字列そのものを出力させています。

モデル マイクロ平均 文化 風習 風土 地理 日本史 行政 法律 医療
sarashina2-8x70b 0.7254 0.7141 0.7750 0.7607 0.6544 0.7843 0.7364 0.6321 0.9167
sarashina2-70b 0.7246 0.7188 0.7450 0.7355 0.6728 0.7638 0.7636 0.6656 0.9167
Llama-3.3-Swallow-70B-v0.4 0.6973 0.6891 0.7750 0.5894 0.5662 0.7755 0.7727 0.7826 0.8542
RakutenAI-2.0-8x7B 0.6327 0.6219 0.7250 0.6171 0.5110 0.7143 0.7091 0.5753 0.8125
plamo-100b 0.6033 0.6016 0.6500 0.6373 0.5037 0.6822 0.6091 0.5151 0.6875
Mixtral-8x7B-v0.1-japanese 0.5929 0.6016 0.6700 0.5793 0.4926 0.6122 0.7364 0.5452 0.6667
Meta-Llama-3.1-405B 0.5712 0.5578 0.5450 0.4836 0.5000 0.6793 0.6455 0.6288 0.6875
llm-jp-3.1-8x13b 0.5682 0.5953 0.6350 0.5819 0.4485 0.5889 0.6273 0.5017 0.6250
Nemotron-4-340B-Base 0.5673 0.5734 0.6150 0.5113 0.4669 0.5948 0.7273 0.5819 0.6667
Qwen2.5-72B 0.5271 0.5219 0.5950 0.4257 0.4375 0.6064 0.6091 0.5619 0.6875

Language

日本語

Dataset Structure

Data Instances

文化カテゴリの問題例は以下の通り:

{
  "qid": "jamcqa-test-culture-00001",
  "category": "culture",
  "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,
}

Data Fields

  • qid (str): 質問を一意識別するための ID
  • category (str): 質問カテゴリ 全8種
    • 文化(culture)、風習(custom)、風土(regional_identity)、地理(geography)、日本史(history)、行政(government)、法律(law)、医療(healthcare)
  • question (str): 質問文
    • カタカナを除き、全角→半角統一済み
    • 内部に改行 \n を含まない
    • 文字列前後の空白は除去済み
  • choice{0..3} (str): 選択肢(choice0〜choice3 の 4つ)
    • カタカナを除き、全角→半角統一済み
    • 内部に改行 \n を含まない
    • 文字列前後の空白は除去済み
  • answer_index (int): choice0choice3 に対応した正解選択肢のインデックス(0-3)

Data Splits

  • dev: カテゴリごとに 4-shot 学習用の 4問を設けています。
  • test: 全 2,309問。

内訳は以下の通り:

カテゴリ dev test
文化(culture) 4 640
風習(custom) 4 200
風土(regional_identity) 4 397
地理(geography) 4 272
日本史(history) 4 343
行政(government) 4 110
法律(law) 4 299
医療(healthcare) 4 48
合計 32 2,309

Licensing Information

How to use

$ python
>>> import datasets
>>> jamcqa = datasets.load_dataset('sbintuitions/JamC-QA', 'v1.0')
>>> print(jamcqa)
DatasetDict({
    test: Dataset({
        features: ['qid', 'category', 'question', 'choice0', 'choice1', 'choice2', 'choice3', 'answer_index'],
        num_rows: 2309
    })
    dev: Dataset({
        features: ['qid', 'category', 'question', 'choice0', 'choice1', 'choice2', 'choice3', 'answer_index'],
        num_rows: 32
    })
})
>>> jamcqa_test = jamcqa['test']
>>> print(jamcqa_test)
Dataset({
    features: ['qid', 'category', 'question', 'choice0', 'choice1', 'choice2', 'choice3', 'answer_index'],
    num_rows: 2309
})
>>> print(jamcqa_test[0])
{'qid': 'jamcqa-test-culture-00001', 'category': 'culture', '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}
>>> 

Citation Information

@inproceedings{Oka2025,
  author={岡 照晃, 柴田 知秀, 吉田 奈央},
  title={JamC-QA: 日本固有の知識を問う多肢選択式質問応答ベンチマークの構築},
  year={2025},
  month={March},
  booktitle={言語処理学会第31回年次大会(NLP2025)},
  pages={839--844},
}