WMTH-100d2exam/epodreczniki

e-podręczniki history textbooks in the data/hf layout

Six Polish history e-textbooks from zpe.gov.pl (231 lessons), with their full text and two kinds of questions:

• generated: about 20–110 exam-style questions per lesson, written by Claude from the lesson text.
• original: the textbook's own exercises (Polecenie / Ćwiczenie).

Questions follow data/hf (format separate-text-and-images-v1), the same way as data/synthetic/rephrased/structured:

• one directory per "exam" with exam.json and key.json;
• a flat items.jsonl with the fields of data/hf/items.jsonl.

file / dir	rows	content
generated/<exam_id>/	538 exams	generated questions, one lesson split into parts of ≤ 48 items (-p1, -p2, …)
original/<exam_id>/	231 exams	the textbook's own exercises for each lesson
items.jsonl	22,129	every question, flat: 19,169 generated + 2,960 original
lessons.jsonl	231	full lesson text in Markdown, with book/chapter metadata; join on lesson_id
books.jsonl	6	full book text (book.md)

Exam id: epod-<book>-<NN>-gen[-pK] / epod-<book>-<NN>-orig. Book codes: starozytnosc, sredniowiecze, nowozytnosc, xix (Długi wiek XIX), xx (ważn@historia, 1918 to the present), wspolna (wspólna historia).

Questions per book:

book	questions
starozytnosc	1,091
sredniowiecze	3,968
nowozytnosc	4,462
xix	4,758
xx	6,686
wspolna	1,164

items.jsonl fields

Same as data/hf/items.jsonl:

• id, exam_id, status, max_points, answer_format, task_type, requirement_ids, sources
• question, source_text, images, image_alt, rubric, key_answer, prompt_text_only

Additional fields:

• lesson: split (generated / original), book, book_title, chapter, lesson_id, lesson_no, lesson, lesson_code, lesson_url.
• generated only: category, level, difficulty, source_format, explanation, evidence, source_question_id.
  • category (12 values) is one of: daty, postacie, miejsca, pojecia, liczby, szczegoly, chronologia, przyczyny_skutki, porownanie, zrodla, podsumowanie, wnioski.
  • level is one of: fakt, zrozumienie, synteza.
  • difficulty is 1–3.
  • evidence is a verbatim quote from the lesson that supports the answer.
• original only: label (e.g. "Ćwiczenie 3"), source_format (the ZPE widget type).

status values:

• synthetic: generated questions.
• original: textbook exercise with an answer key.
• original-no-key: open textbook task with no key. rubric and key_answer are null, as in the rephrased text-only split.

Conversion to hf answer formats

source format	hf answer_format	task_type	key example
single choice	A	closed_choice	C
multiple choice	1: P\n2: F\n3: P (one P/F per option)	true_false	1: P\n2: F\n…
true/false statement	1: P\n2: F\n3: P	true_false	1: F
matching	A: 1\nB: 1	closed_match	A: 2\nB: 1
ordering	open text	closed_order	3, 1, 2 + ordered list
short answer / fill the gap	open text	open_identify	the word(s)
open question	open text	open_explain / open_compare / open_argument / open_source / open_chronology	model answer

• Multiple choice as P/F: the hf enum has no multi-letter format, so multiple choice becomes a P/F list.
• Option order: options are shuffled deterministically. The generator tended to put the correct option first.
• Points: 1 per task. Multi-statement P/F with more than 2 statements and synteza-level open questions are worth 2.
• Rubrics: CKE-style text generated per format.
• Key notes: key.json → notes holds the explanation and the evidence quote.

Original exercises

1,105 of the textbook's interactive exercises were converted into text tasks with keys. Supported types:

• choice;
• ordering and matching;
• crosswords;
• gap texts, with the word bank for drag-and-drop gaps;
• category sorting;
• checkbox and drag-and-drop tables;
• text marking;
• timelines.

Also included: 1,855 open Polecenie tasks without a key.

Not included:

• Image or audio tasks (212): maps, image hotspots, puzzles, picture choice.
• Hotspot illustrations and games (11).
• Empty journal widgets (53).
• Unparsed (13).

Original exercises often refer to the lesson's sources ("według relacji Ibrahima…"). Give the model lessons.jsonl[lesson_id].text as context for them.

Generated questions: method and quality

• Method: one Claude (Opus 5.5) agent per batch of lessons read each lesson in full and wrote the questions by hand, following scripts/epodreczniki_questions_prompt.md.
  • Answers come only from the lesson.
  • No duplicates of the lesson's own tasks.
  • Questions are self-contained.
  • Every section of the lesson is covered.
  • Answer options were balanced across formats and difficulty levels.
• Count: about one question per 250 characters of lesson text, between 20 and 100 per lesson (a few long lessons go up to about 110–147).
• Validation (scripts/epodreczniki_questions.py validate): all 231 lessons pass.
  • schema;
  • per-lesson target count;
  • at least 8 categories and all three levels;
  • no duplicate questions;
  • every evidence quote found verbatim in the lesson.
• Known skew:
  • synteza is about 10–15% per lesson (the target was about 20%), because agents were told not to relabel questions afterwards.
  • Single choice is the largest format.
  • A few early batches (the ważn@historia book and part of Długi wiek XIX) adjusted level labels after writing.
• Errors in the source text: the agents found many factual errors and garbled fragments in the textbooks and did not build questions on them. See ERRATA.md (about 150 lessons listed). Some are serious:
  • the Council of Trent credited with defining papal infallibility;
  • the Kompania Kadrowa given "over 100 thousand" soldiers;
  • Bethlehem placed in Galilee.
• Not reviewed by a human: the questions were checked only by the generating model and by automated validation. A spot check is recommended.
• No curriculum ids: requirement_ids is empty ([]).

Validation (validation.json)

• Relaxed schemas: all 769 exam.json and 769 key.json files pass exam.relaxed.schema.json / key.relaxed.schema.json, copied from data/synthetic/rephrased/structured.
• Strict hf schemas: as expected, no file passes the strict hf schemas, because they describe full 60-point matura papers:
  • max_points must be 60;
  • at least 28 items;
  • at least 1 requirement_ids entry;
  • 3 essay topics.
• Essay placeholder: each key.json has an empty essay (topics: []).

Rebuilding

uv run scripts/epodreczniki_crawler.py --offline          # lesson text from the cache
python3 scripts/epodreczniki_questions.py validate         # question banks
python3 scripts/epodreczniki_questions.py merge            # <book>/questions.jsonl
uv run scripts/build_epodreczniki_dataset.py               # this directory

Licence

Lesson text, sources and exercises come from ZPE e-podręczniki (zpe.gov.pl). Licences vary per resource (often CC BY / CC BY-SA, some "tylko do użytku edukacyjnego"), so check the resource before redistributing. The generated questions are new text derived from the lessons.