exam_sample

Experimental sample dataset: 66 exam questions extracted from 3 Vietnamese exams (one multi-exam source PDF split into 3), with per-question solution text and associated figure images.

Contents

• dataset.jsonl -- 66 rows (22 per exam).
• images/ -- figure images referenced by the samples, namespaced exam-<N>__<original-name> to avoid collisions across exams.
• missing_assets.jsonl -- figures the parser referenced but did not emit as a physical file (see Known issues).
• combined_summary.json -- aggregate audit counts.

Row schema (dataset.jsonl)

{
  "global_id": "exam-1-part_1-q1",
  "source_exam": 1,
  "section_id": "part_1",            // part_1=MCQ, part_2=true/false, part_3=short-answer
  "question_number": "1",
  "question": "... question stem + choices/statements ...",
  "solution": "... worked solution / answer ... (or null)",
  "visualization": [
    {
      "path": "images/exam-1__img_p6_1.png",
      "type": "image",
      "page_number": 7,
      "caption": null,
      "id": "...", "reference": "...",
      "asset_status": "present"       // "present" or "missing"
    }
  ]
}

Each sample has 0..n images. Image paths are relative to the dataset root (images/...).

Composition

• Total: 66 questions (exam 1: 22, exam 2: 22, exam 3: 22).
• Sections: part_1 = 36, part_2 = 12, part_3 = 18.
• Images: 97 present.

Provenance

Generated by an experimental pipeline: split-exams -> preprocess -> line-level JEV classification (labels: question_anchor, question_content, solution, visualization) -> block-export. Question/solution partition is coverage-first and may contain extra solution-side lines.