---
license: apache-2.0
language:
- ja
- en
tags:
- medical-qa
- multiple-choice
- japanese-medical-exam
- curated-subset
- difficult-samples
- text-only
size_categories:
- n<1K
task_categories:
- question-answering
- multiple-choice
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: question
    dtype: string
  - name: options
    sequence: string
  - name: answer_idx
    dtype: int64
  - name: correct_answer
    dtype: string
  - name: sample_id
    dtype: string
  - name: year
    dtype: string
  - name: text_only
    dtype: bool
  - name: points
    dtype: int64
  - name: difficulty_score
    dtype: float64
  - name: model_failures
    dtype: int64
  - name: failure_type
    dtype: string
  splits:
  - name: train
    num_bytes: 114138
    num_examples: 146
  download_size: 66955
  dataset_size: 114138
---

# IgakuQA Curated Subset (Text-Only)

This dataset contains 66 carefully selected **text-only** samples from the IgakuQA medical exam dataset, 
curated using advanced difficulty assessment and model evaluation techniques.

## Dataset Description

This is a high-quality subset of the IgakuQA Japanese medical exam dataset, selected based on:
- **Difficulty Score**: Measures how challenging the sample is for AI models
- **Consistency Score**: Evaluates response consistency across different models
- **Model Agreement**: Measures consensus among different AI models
- **Response Quality**: Assesses the quality of model responses
- **Text-Only Constraint**: Only includes questions that don't require images

## Recent Update

**🔄 Dataset Cleaned**: Filtered to only include text-only samples
- **Original**: 100 samples (mixed text-only and image-based)
- **Updated**: 66 samples (text-only only)
- **Removed**: 34 image-based samples

This ensures compatibility with language models that don't support image inputs via agent services.

## Selection Methodology

The samples were selected using a sophisticated sample picker that:
1. Evaluates samples using multiple state-of-the-art language models
2. Calculates difficulty metrics based on model performance
3. Ranks samples by their educational and evaluation value
4. Selects the most challenging and informative samples
5. **Filters to text-only samples** for broader compatibility

## Dataset Structure

### Fields

**Core QA Fields:**
- `question`: The medical exam question (in Japanese)
- `options`: List of answer options
- `correct_answer`: The correct answer text
- `answer_idx`: Index of the correct answer (0-based)

**Metadata:**
- `sample_id`: Original sample ID from IgakuQA
- `year`: Exam year
- `text_only`: Whether the question is text-only (always `true` in this dataset)
- `points`: Point value of the question

**Selection Context:**
- `difficulty_score`: How difficult this sample is for models (higher = more difficult)
- `picker_sample_id`: Sample ID assigned by the picker tool

## Usage

This curated subset is ideal for:
- Medical AI model evaluation and benchmarking with text-only models
- Fine-tuning on challenging Japanese medical questions
- Agent service evaluation where image support is limited
- Research on medical knowledge capabilities without visual components
- Educational purposes for understanding model limitations in medical domains

## Language

The questions are primarily in Japanese, as they are from Japanese medical licensing examinations.

## Compatibility

- ✅ **Text-only models** (GPT, Claude, Gemini via API)
- ✅ **Agent services** without image support
- ✅ **Fine-tuning pipelines** focused on text understanding
- ❌ **Image-based medical reasoning** (use full IgakuQA for that)

## Original Dataset

This subset is derived from the IgakuQA dataset - a collection of Japanese medical exam questions.

## Citation

If you use this curated subset, please cite both this dataset and the original IgakuQA dataset.