#!/usr/bin/env python3 """Parse EDN France exam PDFs (CNG annales 2023-2025) into question/answer CSVs. Reads pdftotext -layout text dumps, splits stems/questions, extracts choices, links embedded figures, writes one CSV per year. No answer key exists (official CNG PDFs are questions-only; community corrections are platform-locked) -> correct_answer left empty. """ import csv, os, re, sys try: import fitz HAVE_FITZ = True except ImportError: HAVE_FITZ = False SRC = os.path.expanduser('~/workspace/medical-ai-leaderboard-staging/repo/data/updates_raw/edn_france') TXTDIR = '/tmp/edn_txt' OUTDIR = os.path.expanduser('~/workspace/medical-ai-leaderboard-staging/repo/data/processed_csv') IMGDIR = os.path.join(OUTDIR, 'images') import subprocess as _sp # (year, paper_label, pdf_rel, txt_name) PAPERS_TXT = [ ('2023', 'P1', '2023/EDN_P1.pdf', '2023_EDN_P1.txt'), ('2023', 'P2', '2023/EDN_P2.pdf', '2023_EDN_P2.txt'), ('2023', 'P3', '2023/EDN_P3.pdf', '2023_EDN_P3.txt'), ('2023', 'LCA', '2023/lca/EDN_LCA_P.pdf', '2023_LCA.txt'), ('2024', 'P1', '2024/EDN_P1.pdf', '2024_EDN_P1.txt'), ('2024', 'P2', '2024/EDN_P2.pdf', '2024_EDN_P2.txt'), ('2024', 'P3', '2024/EDN_P3.pdf', '2024_EDN_P3.txt'), ('2024', 'LCA', '2024/lca/EDN_LCA_P.pdf', '2024_LCA.txt'), ('2025', 'R1', '2025/EDN_P1_reserve.pdf', '2025_EDN_P1.txt'), ('2025', 'P2', '2025/EDN_P2.pdf', '2025_EDN_P2.txt'), ('2025', 'P3', '2025/EDN_P3.pdf', '2025_EDN_P3.txt'), ('2025', 'LCA', '2025/lca/EDN_LCA_P.pdf', '2025_LCA.txt'), ] def ensure_text(): os.makedirs(TXTDIR, exist_ok=True) for year, paper, pdf_rel, txt_name in PAPERS_TXT: dest = os.path.join(TXTDIR, txt_name) if not os.path.exists(dest): _sp.run(['pdftotext', '-layout', os.path.join(SRC, pdf_rel), dest], check=True) # (year, paper_label, pdf_name) PAPERS = [ ('2023', 'P1', 'EDN_P1.pdf'), ('2023', 'P2', 'EDN_P2.pdf'), ('2023', 'P3', 'EDN_P3.pdf'), ('2023', 'LCA', 'lca/EDN_LCA_P.pdf'), ('2024', 'P1', 'EDN_P1.pdf'), ('2024', 'P2', 'EDN_P2.pdf'), ('2024', 'P3', 'EDN_P3.pdf'), ('2024', 'LCA', 'lca/EDN_LCA_P.pdf'), ('2025', 'R1', 'EDN_P1_reserve.pdf'), ('2025', 'P2', 'EDN_P2.pdf'), ('2025', 'P3', 'EDN_P3.pdf'), ('2025', 'LCA', 'lca/EDN_LCA_P.pdf'), ] PDFMAP = {(y, p): f'{y}/{pdf}' for y, p, pdf in PAPERS} MIN_DIM = 100 # min(w,h) for a figure worth extracting (<100 = UI decoration) RE_STEM = re.compile(r'^(KFP|DP|QI|TCS|LCA) (\d+)\s*$') RE_Q = re.compile(r'^Question (\d+) - (.+?)\s*$') RE_FOOTER = re.compile(r'^Epreuve .*?\b(\d+)/(\d+)\s*$') RE_TITLE = re.compile(r'^\s*Epreuve de .*$') RE_PROP = re.compile(r'^ {1,12}Proposition ([A-P])\s*$') RE_REPONSE = re.compile(r'^\s*Réponse\s*:\s*$') RE_FIN = re.compile(r'--- Fin de copie ---') RE_ZONELABEL = re.compile(r'^\s*(\d+)\s*-\s*(.*?)\s*$') LIKERT = ['Très peu probable', 'Peu probable', 'Ni plus ni moins', 'Probable', 'Très probable'] IMG_KW = re.compile( r'ci-dessous|ci-dessus|' r'\b(figures?|images?|clichés?|photos?|schémas?|tracés?|courbes?)\b|' r'\b(ECG|électrocardiogrammes?)\b|' r'\b(radiographies?|scanners?|scanographies?|coupes?)\b|' r'\bradio\b', re.I) issues = [] # (year, paper, qid, message) def clean(s): s = re.sub(r'\s+', ' ', s).strip() return s def join_lines(lines): """Join wrapped layout lines; trailing '-' = hyphenation -> concat directly.""" out = '' for ln in lines: ln = ln.strip() if not ln: continue if out.endswith('-'): out += ln else: out = (out + ' ' + ln).strip() if out else ln return clean(out) def parse_paper(year, paper, txt_path): raw = open(txt_path, encoding='utf-8').read().replace('\f', '\n') lines = raw.split('\n') questions = [] page = 1 stem = None # (type, num) stem_text_lines = [] # vignette text between stem header and its first question cur = None # current question dict q_pages = [] def flush(): nonlocal cur if cur is not None: finalize(cur) questions.append(cur) cur = None def finalize(q): # split body into question text vs choices body = q['body'] # list of (line, page) text_lines, choices, zone_labels = [], [], [] mode = 'text' cur_letter, cur_clines = None, [] likert_done = False in_legend = False # QZP: inside figure zone-label legend for ln, pg in body: if RE_FIN.search(ln): break # QZP figure legend: "N - label" lines (label may wrap onto following lines) if q['is_qzp'] and mode == 'text': mz = RE_ZONELABEL.match(ln) if mz: zone_labels.append([mz.group(1), clean(mz.group(2))]) in_legend = True continue if in_legend and ln.strip(): zone_labels[-1][1] = (zone_labels[-1][1] + ' ' + clean(ln)).strip() continue mp = RE_PROP.match(ln) if mp and mode == 'text': # start of choices if cur_letter: choices.append((cur_letter, join_lines(cur_clines))) cur_letter, cur_clines = mp.group(1), [] mode = 'choices' continue if mode == 'text' and RE_REPONSE.match(ln): mode = 'reponse' # QROC answer area: ignore rest continue if mode == 'text' and 'Très peu probable' in ln and 'Très probable' in ln: parts = [p.strip() for p in re.split(r'\s{2,}', ln.strip()) if p.strip()] if len(parts) >= 5: choices.extend([(chr(65 + i), parts[i]) for i in range(5)]) else: # fallback: fixed labels choices.extend([(chr(65 + i), LIKERT[i]) for i in range(5)]) likert_done = True mode = 'likert' continue if mode == 'choices': if mp: # next proposition choices.append((cur_letter, join_lines(cur_clines))) cur_letter, cur_clines = mp.group(1), [] else: cur_clines.append(ln) continue if mode == 'reponse': continue if mode == 'likert': continue text_lines.append(ln) if cur_letter: choices.append((cur_letter, join_lines(cur_clines))) q['choices'] = choices q['text'] = join_lines(text_lines) if zone_labels: zl = '; '.join(f'{n} - {t}' for n, t in zone_labels) q['text'] = (q['text'] + f' [Zones de la figure : {zl}]').strip() q['pages'] = sorted(set(q_pages)) for ln in lines: if RE_FIN.search(ln): flush() break mf = RE_FOOTER.match(ln) if mf: page = int(mf.group(1)) continue if RE_TITLE.match(ln): continue ms = RE_STEM.match(ln) if ms: flush() stem = (ms.group(1), ms.group(2)) stem_text_lines = [] continue mq = RE_Q.match(ln) if mq: flush() fmt = mq.group(2) cur = {'sub': int(mq.group(1)), 'format': fmt, 'body': [], 'stem': stem, 'start_page': page, 'stem_text': join_lines(stem_text_lines), 'is_qzp': 'zone à pointer' in fmt} q_pages = [page] continue if cur is not None: if ln.strip() == '': cur['body'].append(('', page)) else: cur['body'].append((ln, page)) q_pages.append(page) elif stem is not None and ln.strip(): stem_text_lines.append(ln) flush() # assign ids for q in questions: st, sn = q['stem'] if q['stem'] else ('NA', '0') q['qid'] = f"edn{year}_{paper}_{st}{sn}_q{q['sub']}" return questions def extract_images(year, paper, pdf_rel): """Save embedded figures (min dim >= MIN_DIM, skip title page). Returns {1-idx page: [fnames]}.""" page_map, saved = {}, {} if not HAVE_FITZ: return page_map, 0 d = fitz.open(os.path.join(SRC, pdf_rel)) outdir = os.path.join(IMGDIR, f'edn_france_{year}') os.makedirs(outdir, exist_ok=True) for pno, page in enumerate(d): if pno == 0: continue # title page: EDN banner for img in page.get_images(full=True): xref, w, h = img[0], img[2], img[3] if min(w, h) < MIN_DIM: continue if xref not in saved: fname = f'edn_france_{year}_{paper}_p{pno + 1:03d}_xref{xref}.png' try: pix = fitz.Pixmap(d, xref) if pix.n > 4: pix = fitz.Pixmap(fitz.csRGB, pix) # skip blank/uniform images (masks, empty placeholders) s = pix.samples step = max(1, len(s) // 4000) vals = s[::step] if max(vals) - min(vals) < 12: continue pix.save(os.path.join(outdir, fname)) except Exception as e: issues.append((year, paper, f'image save failed {fname}: {e}')) continue saved[xref] = fname page_map.setdefault(pno + 1, []) if saved[xref] not in page_map[pno + 1]: page_map[pno + 1].append(saved[xref]) d.close() return page_map, len(saved) def link_images(year, paper, pdf_rel, qs): page_map, n_saved = extract_images(year, paper, pdf_rel) linked = 0 claimed = {} # image fname -> stem key of question that claimed it for q in qs: cands = [] for pg in q['pages'] + [max(q['pages']) + 1]: cands.extend(page_map.get(pg, [])) cands = list(dict.fromkeys(cands)) needs = q['is_qzp'] or IMG_KW.search(q['text'] or '') or IMG_KW.search(q.get('stem_text') or '') if needs: stemkey = q['stem'] # prefer a figure not already claimed by a *different* stem # (sub-questions of one stem legitimately share a figure) pick = None for c in cands: if claimed.get(c, stemkey) == stemkey: pick = c break if pick is None and cands: pick = cands[0] if pick: q['image_file'] = pick claimed.setdefault(pick, stemkey) linked += 1 else: issues.append((year, q['qid'], 'QZP but no figure found' if q['is_qzp'] else 'image keyword but no figure found')) else: q['image_file'] = '' return n_saved, linked def main(): ensure_text() all_q = {} img_stats = {} for year, paper, pdf_rel, txt_name in PAPERS_TXT: txt = os.path.join(TXTDIR, txt_name) qs = parse_paper(year, paper, txt) n_saved, linked = link_images(year, paper, pdf_rel, qs) img_stats[(year, paper)] = (n_saved, linked) all_q.setdefault(year, []).extend(qs) print(f'{year} {paper}: {len(qs)} questions, {n_saved} images, {linked} linked') # write CSVs cols = ['exam_id', 'question_number', 'question_id', 'question_text', 'choice_a', 'choice_b', 'choice_c', 'choice_d', 'choice_e', 'correct_answer', 'image_file', 'language', 'country', 'year'] for year in ('2023', '2024', '2025'): qs = all_q[year] path = os.path.join(OUTDIR, f'edn_france_{year}.csv') with open(path, 'w', encoding='utf-8', newline='') as fh: w = csv.writer(fh) w.writerow(cols) for i, q in enumerate(qs, 1): ch = {letter: text for letter, text in q['choices']} if len(q['choices']) > 5: issues.append((year, q['qid'], f"{len(q['choices'])} proposals; kept A-E, dropped {len(q['choices'])-5}")) if not q['choices'] and not q['is_qzp'] and 'ouverte' not in q['format'] and 'concordance' not in q['format']: issues.append((year, q['qid'], f"no choices parsed (format: {q['format']})")) if not q['text']: issues.append((year, q['qid'], 'empty question text')) w.writerow([ f'edn_france_{year}', i, q['qid'], q['text'], ch.get('A', ''), ch.get('B', ''), ch.get('C', ''), ch.get('D', ''), ch.get('E', ''), '', # correct_answer: no key exists q.get('image_file', ''), 'fr', 'France', year, ]) print(f'wrote {path}: {len(qs)} rows') print('\nISSUES:', len(issues)) for it in issues[:60]: print(it) with open(os.path.join(OUTDIR, 'edn_france_build_issues.log'), 'w', encoding='utf-8') as fh: fh.write(f'total issues: {len(issues)}\n') for it in issues: fh.write(repr(it) + '\n') print('full issue list -> edn_france_build_issues.log') if __name__ == '__main__': main()