AugeoBench

AugeoBench is a multimodal QA benchmark consisting of Japanese entrance-exam-style geometry questions. Each instance provides a contextual description, a question text, a reference image, and a correct answer.

Dataset Structure

• id: Unique ID for the question
• url: URL of the question
• question_context_ja: Context of the question
• question_text_ja: Question text
• question_information_ja: Additional information about the question
• answer_exact: Exact answer
• answer_text_ja: Answer text
• question_image: Question image
• answer_image: Answer image
• Genre: Genre of the question
• Remarks: Remarks about the question

Example

{
  "id": "1",
  "url": "https://sansu-seijin.jp/nyushimondai/2025-nyushimondai/19331/",
  "question_context_ja": "図において,MはBEの真ん中の点であり...",
  "question_text_ja": "線分ADと線分BCの関係を答えなさい。",
  "question_information_ja": "図において,MはBEの真ん中の点であり...",
  "answer_exact": "平行",
  "answer_text_ja": "視覚化した図において,同じ色のついた三角形は角度が等しい相似な図形です。\\\\(\\angle ADC = \\angle BCD\\\\) で錯角が等しいので,線分ADと線分BCは平行です。",
  "question_image": PIL.Image.Image,
  "answer_image": PIL.Image.Image,
  "Genre": "Relation, Focus",
  "Remarks": "placeholder_remarks"
}

Updating procedure

1. Annotate the dataset and save it as AugeoBench_annotationsheet.xlsx
2. Cleanse the dataset with uv run python cleanse.py
3. Zip the images with bash prepare.sh in top-level directory
4. Run uv run python -m augeobench.push_to_hub