JMedQA / README.md
YYama0's picture
Update README.md
602344b verified
|
Raw
History Blame Contribute Delete
19.7 kB
metadata
license: other
language:
  - ja
task_categories:
  - question-answering
  - multiple-choice
tags:
  - medical
  - japanese
  - exam
  - benchmark
  - jmedqa
  - multimodal
  - vision-language
pretty_name: JMedQA
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: benchmark
        path: jmedqa.csv

JMedQA: Benchmarking Large Language Models and Vision-Language Models on the Japanese Medical Licensing Examination

JMedQA is a Japanese medical question-answering benchmark derived from Japan's National Medical Examination materials publicly released by the Ministry of Health, Labour and Welfare (MHLW).

The dataset supports both text-only large language model (LLM) evaluation and vision-language model (VLM) evaluation using associated examination images.

Its image-dependency annotations allow researchers to construct different evaluation subsets, including:

  • text-only questions that can be answered without images;
  • questions that include images but remain answerable from text alone;
  • questions that require visual information;
  • image-only or image-centered questions; and
  • paired text-only and multimodal settings for analyzing the contribution of visual input.

The current release contains structured question data and associated cropped examination images.

  • File: jmedqa.csv
  • Rows: 3581
  • Target exam years: 2018-2026
  • Current modality: structured text data and associated images
  • Image path references: 1326
  • Unique referenced image files: 1258

What Is Included

Each row represents a Japanese medical examination question, consisting primarily of multiple-choice items and a small number of numeric-answer items, together with evaluation-oriented metadata.

Main columns:

  • id: sequential row ID
  • problem_unique_id: exam-style unique ID, e.g. 2018A001
  • year: exam year
  • section: block/section, A-F
  • clinical_area: medical domain label
  • question: question text used for text-only evaluation
  • question_raw: original question text retaining explicit image-reference phrases
  • options_json: options as a JSON object, typically keys a-e
  • answer_json: gold answer as a JSON array
  • answer_mode: option or numeric
  • answer_count: expected number of answers
  • is_calc: calculation-related flag inherited from the source data
  • is_linked: linked-question flag
  • is_case_based: case-based question flag
  • image: image-reference field
  • image_paths_json: JSON mapping from image references to relative image paths
  • has_unpublished_image: whether the row includes an MHLW unavailable-image notice
  • unpublished_image_refs_json: image references corresponding to unavailable-image notices
  • image_dependency: annotation describing how strongly the question depends on visual information

question and question_raw

The dataset provides two versions of the question text to support different evaluation settings.

  • If image_dependency=none, question is generally identical to question_raw.
  • If image_dependency != none, question is generated from question_raw by removing explicit image-reference phrases.
  • question supports text-only evaluation without presenting image-reference phrases to the model.
  • question_raw preserves the original image references and is intended for multimodal evaluation with the corresponding images.

Because the transformation follows a deletion-only policy, a subset of the resulting question texts may contain slightly unnatural phrasing.

Images

Associated images are included as cropped PNG files under relative paths such as:

images/2026/A/2026A_No01A.png

The image_paths_json column maps each original image reference to one or more image paths.

Example:

{"別冊No.1A": ["images/2026/A/2026A_No01A.png"]}

Some images are not publicly shown in the original MHLW materials. In such cases, the corresponding notice image from the source material is cropped and included as-is, and the row is flagged with:

  • has_unpublished_image=true
  • unpublished_image_refs_json

image_dependency Definitions

  • none: no image dependency
  • enough text: an image is referenced, but the textual information is sufficient to answer the question
  • not enough text: an image is referenced and the textual information alone is insufficient
  • image only: the question text consists primarily of an image reference
  • image question: the question explicitly asks about visual findings, interpretation, or diagnosis

Dataset Profile

  • answer_mode=option: 3556
  • answer_mode=numeric: 25
  • is_case_based=true: 2251
  • is_linked=true: 446
  • has_unpublished_image=true: 18
  • rows with image references: 944
  • image path references: 1326
  • unique referenced image files: 1258

image_dependency distribution:

  • none: 2608 (72.83%)
  • enough text: 670 (18.71%)
  • not enough text: 172 (4.80%)
  • image only: 7 (0.20%)
  • image question: 124 (3.46%)

Year distribution:

  • 2018: 399
  • 2019: 395
  • 2020: 398
  • 2021: 400
  • 2022: 396
  • 2023: 394
  • 2024: 399
  • 2025: 400
  • 2026: 400

Intended Evaluation Settings

JMedQA is designed to support several complementary evaluation settings.

1. Text-Only Evaluation

For evaluation of text-only LLMs, we recommend using questions with:

  • image_dependency=none; or
  • image_dependency in {none, enough text} when a broader text-answerable subset is desired.

The question field should be used as the model input in this setting.

2. Multimodal Evaluation

For VLM evaluation, use question_raw together with the images listed in image_paths_json.

The multimodal subset can include all questions for which associated images are available, with results reported separately by image_dependency.

3. Modality-Gap Analysis

Questions with both usable text and associated images can be evaluated under two conditions:

  • text-only input; and
  • text-plus-image input.

This setting enables analysis of whether and to what extent visual information improves medical question-answering performance.

Results should be reported separately by evaluation setting and image-dependency category to ensure comparability.

Data Preparation Note

This release was created from PDFs publicly provided by Japan's Ministry of Health, Labour and Welfare. Problem extraction and data preparation used Vertex AI and Codex (GPT-5.5), followed by text cleanup, metadata annotation, and quality checks.

The preparation process aimed to preserve the original wording. Quality checks were performed on a small scale, but a small number of extraction, annotation, or formatting errors may remain.

For image-related rows, explicit image-reference phrases were removed from the question field using a deletion-only policy. The corresponding references are retained in question_raw, and image file paths are stored in image_paths_json.

Evaluation-Oriented Preprocessing

To support single-question evaluation workflows:

  • shared case descriptions from linked question groups were propagated so that each item can be evaluated independently;
  • questions excluded from official-style scoring were removed from this release;
  • explicit image-reference phrases were removed from question, while being retained in question_raw;
  • numeric targets are stored as strings in JSON arrays in answer_json;
  • cropped image files are referenced by relative paths in image_paths_json; and
  • notice images for source images that are not publicly shown are retained as-is and explicitly flagged.

Quality Control

The current release was checked using a combination of automated validation and small-scale manual review. These checks covered structured fields, image path consistency, and selected visual samples. However, the dataset is derived from OCR and automated processing, so a small number of residual errors may remain.

License

The original examination content is derived from the Japanese National Medical Examination materials published by Japan's Ministry of Health, Labour and Welfare (MHLW). Please use the dataset and associated images in accordance with the applicable MHLW terms of use. Source: https://www.mhlw.go.jp/

Quick Load Example (Python)

from datasets import load_dataset

dataset = load_dataset(
    "SIP-med-LLM/JMedQA",
    split="benchmark",
)

print(dataset)
print(dataset[0])

Roadmap / Planned Updates

JMedQA is actively maintained. Planned future work includes:

  • providing standardized text-only and multimodal evaluation subsets;
  • releasing evaluation and scoring code;
  • adding newer examination years;
  • further answer verification and quality control; and
  • publishing benchmark results across a wide range of LLMs and VLMs.

Citation

If you use JMedQA in your research, please cite the dataset as follows:

@misc{yamagishi2026jmedqa,
  title        = {JMedQA: Benchmarking Large Language Models and Vision-Language Models on the Japanese Medical Licensing Examination},
  author       = {Yamagishi, Yosuke and Kobayashi, Kazuma and Shibaki, Ryota and Aizawa, Akiko and Kurohashi, Sadao},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/SIP-med-LLM/JMedQA}},
  note         = {Hugging Face dataset},
  urldate      = {2026-06-25}
}

JMedQA:日本の医師国家試験を用いた大規模言語モデル・視覚言語モデルのベンチマーク評価

JMedQAは、厚生労働省が公開している日本の医師国家試験資料をもとに構築した、日本語医学QAベンチマークです。

本データセットは、テキストのみを入力とする大規模言語モデル(LLM)の評価に加え、関連画像を用いた視覚言語モデル(VLM)の評価にも対応することを目的として設計されています。

image_dependency アノテーションを利用することで、以下のような評価サブセットを構成できます。

  • 画像を必要としないテキスト問題
  • 画像を参照するものの、テキストのみでも解答可能な問題
  • 解答に視覚情報を必要とする問題
  • 画像のみ、または画像所見を中心とする問題
  • 同一問題を画像なし/画像ありで評価し、視覚情報の寄与を分析する設定

現在の公開版には、構造化された問題データと対応する別冊画像が含まれています。

  • ファイル:jmedqa.csv
  • 件数:3581
  • 対象年度:2018-2026
  • 現在の公開内容:構造化テキストデータおよび関連画像
  • 画像パス参照数:1326
  • ユニーク画像ファイル数:1258

含まれる内容

各行は、日本の医師国家試験の設問と評価用メタデータから構成されます。大部分は選択肢問題であり、一部に数値回答問題が含まれます。

主なカラム:

  • id:行ID
  • problem_unique_id:問題固有ID(例:2018A001
  • year:試験年度
  • section:ブロック(AF
  • clinical_area:医学領域ラベル
  • question:テキストのみの評価に使用する問題文
  • question_raw:明示的な画像参照表現を保持した元の問題文
  • options_json:選択肢を格納したJSONオブジェクト(通常は ae
  • answer_json:正解を格納したJSON配列
  • answer_modeoption または numeric
  • answer_count:想定される正答数
  • is_calc:元データに由来する計算問題フラグ
  • is_linked:連問フラグ
  • is_case_based:症例ベース問題フラグ
  • image:画像参照フィールド
  • image_paths_json:画像参照から相対画像パスへの対応を格納したJSON
  • has_unpublished_image:厚労省公開資料上で画像非公開の注記画像を含むかどうか
  • unpublished_image_refs_json:非公開注記画像に該当する別冊参照のJSON配列
  • image_dependency:設問の視覚情報への依存度を表すラベル

questionquestion_raw

テキストのみの評価とマルチモーダル評価の両方に対応するため、2種類の問題文を収録しています。

  • image_dependency=none の場合、question は原則として question_raw と同一です。
  • image_dependency != none の場合、question_raw から明示的な画像参照表現のみを削除して question を作成しています。
  • question はテキストのみの評価に使用することを想定しています。
  • question_raw は画像への参照を保持しており、対応画像と組み合わせたマルチモーダル評価に使用することを想定しています。

この変換は削除のみを行う方針(deletion-only policy)で実施しているため、一部の question では表現がやや不自然になっています。

画像

関連画像は、以下のような相対パスのPNGファイルとして含まれます。

images/2026/A/2026A_No01A.png

image_paths_json には、元の別冊参照と画像パスの対応が格納されています。

例:

{"別冊No.1A": ["images/2026/A/2026A_No01A.png"]}

一部の画像は、元の厚労省公開資料上で公開されていません。その場合は、元資料に含まれる非公開注記画像をそのままクロップして収録し、以下の列で明示しています。

  • has_unpublished_image=true
  • unpublished_image_refs_json

image_dependency の定義

  • none:画像依存なし
  • enough text:画像参照はあるが、本文テキストのみで解答可能
  • not enough text:画像参照があり、本文テキストだけでは情報が不足
  • image only:問題文の大部分が画像参照で、テキスト情報がほとんどない
  • image question:画像所見、画像診断、視覚的な解釈などを直接問う問題

現在のファイル統計

  • answer_mode=option3556
  • answer_mode=numeric25
  • is_case_based=true2251
  • is_linked=true446
  • has_unpublished_image=true18
  • 画像参照を持つ行:944
  • 画像パス参照数:1326
  • ユニーク画像ファイル数:1258

image_dependency の分布:

  • none260872.83%
  • enough text67018.71%
  • not enough text1724.80%
  • image only70.20%
  • image question1243.46%

年度別件数:

  • 2018399
  • 2019395
  • 2020398
  • 2021400
  • 2022396
  • 2023394
  • 2024399
  • 2025400
  • 2026400

想定する評価設定

JMedQAは、相互に補完的な複数の評価設定に対応するよう設計されています。

1. テキストのみの評価

テキストのみを入力とするLLMの評価では、原則として以下の問題を使用することを推奨します。

  • 厳密なテキストオンリー設定:image_dependency=none
  • より広いテキスト解答可能設定:image_dependencynone または enough text

この評価では、question フィールドをモデルへの入力として使用します。

2. マルチモーダル評価

VLM評価では、question_rawimage_paths_json に記録された画像を組み合わせて使用します。

画像が利用可能な全問題を対象とするほか、image_dependency ごとに性能を分けて報告することで、視覚情報への依存度に応じた評価が可能になります。

3. モダリティ差の分析

テキストと画像の両方が利用できる問題について、以下の2条件を比較できます。

  • テキストのみを入力した場合
  • テキストと画像を入力した場合

これにより、視覚情報が医学QA性能をどの程度改善するかを分析できます。

評価結果を報告する際は、使用した評価設定と image_dependency の対象カテゴリを明記することを推奨します。

データ作成に関する補足

本リリースは、厚生労働省が公開しているPDFをもとに作成しました。問題抽出およびデータ作成には Vertex AI と Codex(GPT-5.5)を用い、その後、表記整備、メタデータ付与、品質確認を行いました。

原文の表現を保持することを意図しています。品質確認は小規模に実施していますが、OCRおよび自動処理に由来する少数の抽出ミス、注釈ミス、表記上の不備が残っている可能性があります。

画像関連行については、明示的な画像参照表現を question から削除しています。この処理は削除のみを行う方針で実施しているため、一部の問題文がやや不自然になっています。画像参照を含む元の表現は question_raw に保持し、対応画像へのパスは image_paths_json に格納しています。

評価向けの前処理

1問単位での推論・採点を可能にするため、以下の前処理を実施しています。

  • 連問で共有される症例情報を各設問に付与し、単独で評価できる形式に変換
  • 公式採点の対象外となる問題を公開版から除外
  • 明示的な画像参照表現を question から削除し、question_raw には保持
  • 数値回答問題の正解を、文字列からなるJSON配列として answer_json に格納
  • クロップ済み画像ファイルを image_paths_json から参照可能な相対パスで収録
  • 元資料上で公開されていない画像については、非公開注記画像をそのまま保持し、該当行を明示的にフラグ付け

品質確認

現在のリリースでは、自動検証と小規模な手動確認を組み合わせて品質確認を行っています。構造化フィールド、画像パスの整合性、および一部の画像内容を確認していますが、OCRおよび自動処理に基づくデータであるため、少数の不備が残っている可能性があります。

ライセンス

原典の試験内容は、厚生労働省が公開する日本の医師国家試験資料に由来します。本データセットおよび関連画像の利用にあたっては、適用される厚生労働省の利用条件に従ってください。出典:https://www.mhlw.go.jp/

読み込み例(Python)

from datasets import load_dataset

dataset = load_dataset(
    "SIP-med-LLM/JMedQA",
    split="benchmark",
)

print(dataset)
print(dataset[0])

今後の予定

JMedQAは継続的に整備しており、以下を予定しています。

  • 標準化されたテキストオンリーおよびマルチモーダル評価サブセットの提供
  • 評価・採点コードの公開
  • 新しい試験年度の追加
  • 正答検証および品質管理の強化
  • 多数のLLMおよびVLMによるベンチマーク結果の公開

引用

JMedQAを研究で利用した場合は、以下を引用してください。

@misc{yamagishi2026jmedqa,
  title        = {JMedQA: Benchmarking Large Language Models and Vision-Language Models on the Japanese Medical Licensing Examination},
  author       = {Yamagishi, Yosuke and Kobayashi, Kazuma and Shibaki, Ryota and Aizawa, Akiko and Kurohashi, Sadao},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/SIP-med-LLM/JMedQA}},
  note         = {Hugging Face dataset},
  urldate      = {2026-06-25}
}