Files
Speech/tools/course_bank/variants.py
T
Eugenius Naumenko f3d42ee2b6 Mehrsprachige Bedienung und versionierte Aufgabenvarianten einführen
Übersetzungskataloge, getrennte Sprachpräferenzen und einen versionierten Aufgabenpilot ergänzen. Historische Aufgaben und Bewertungen bleiben erhalten; äquivalente Ansichten teilen Nachweise. Lokal mit Race-Tests und PostgreSQL, Vet, Web-Build, Bank-Audits und Browser-Fixtures geprüft. Dieser Stand dient als Ausgangspunkt für die anschließend beauftragte Zusammenführung zur Muttersprache.
2026-11-07 16:26:26 +00:00

104 lines
5.8 KiB
Python

"""Offline build/audit of explicitly reviewed explanation-language variants."""
import copy
import hashlib
import json
import pathlib
BASE = pathlib.Path(__file__).resolve().parent
DEST = BASE.parents[1] / 'apps/api/internal/store/coursebank/variants'
BANK = DEST.parent / 'en-US/a1-v2'
def checked_sources():
review_bytes = (BASE / 'variant_review.json').read_bytes()
review = json.loads(review_bytes)
assert review['reviewed'] and review['entry_count'] == 12
for name, digest in review['sources'].items():
assert hashlib.sha256((BASE / name).read_bytes()).hexdigest() == digest, f'Variant review is stale: {name}'
return json.loads((BASE / 'variant_sources.json').read_text()), review_bytes
def compile_variants():
source, review_bytes = checked_sources()
entries = []
for row in source['entries']:
candidates = json.loads((BANK / f'introductions.{row["skill"]}.json').read_text())
base = candidates[row['number'] - 1]
task = copy.deepcopy(base['task'])
variant_id = f'intro_en_explanations_v1_{row["skill"]}_{row["number"]:02d}'
canonical = base['id'] if row['equivalence'] == 'translation' else variant_id
task.update(instruction_language=source['instruction_language'], variant_id=variant_id,
canonical_task_id=canonical, content_revision=source['revision'], instruction=row['instruction'])
task['prompt'] = row['instruction']
if 'stimulus' in row:
task['stimulus'] = row['stimulus']
task['prompt'] += ' ' + row['stimulus']['text']
key = task['answer_key']
key['explanation'] = row['explanation']
key['rubric'] = row.get('rubric', ['Choose the one answer supported by the sentence, word meaning or supplied text.'])
key['reference_answer_language'] = source['instruction_language'] if task['skill'] == 'vocabulary' else task['target_language']
for option in task.get('options', []):
option['role'] = 'meaning' if task['skill'] == 'vocabulary' else 'target'
option['language'] = source['instruction_language'] if option['role'] == 'meaning' else task['target_language']
if 'option_meanings' in row:
option['label'] = row['option_meanings'][option['label']]
if option['id'] == key.get('correct_option_id'):
key['reference_answer'] = option['label']
entries.append(dict(id=variant_id, revision=source['revision'], base_entry_id=base['id'],
canonical_task_id=canonical, instruction_language=source['instruction_language'],
module_title=source['module_title'], module_description=source['module_description'],
equivalence=row['equivalence'], reviewed=True, task=task))
return entries, review_bytes
def build():
entries, review_bytes = compile_variants()
DEST.mkdir(parents=True, exist_ok=True)
data = (json.dumps(entries, ensure_ascii=False, indent=2) + '\n').encode()
name = 'introductions.en-US.v1.json'
(DEST / name).write_bytes(data)
manifest = dict(total=len(entries), editorial_review_sha256=hashlib.sha256(review_bytes).hexdigest(),
files={name: hashlib.sha256(data).hexdigest()})
(DEST / 'manifest.json').write_text(json.dumps(manifest, indent=2) + '\n')
print('Wrote 12 reviewed English-explanation pilot variants across six learning areas.')
def audit():
expected, review_bytes = compile_variants()
manifest = json.loads((DEST / 'manifest.json').read_text())
assert manifest['total'] == 12
assert manifest['editorial_review_sha256'] == hashlib.sha256(review_bytes).hexdigest()
actual = []
for name, digest in manifest['files'].items():
data = (DEST / name).read_bytes()
assert hashlib.sha256(data).hexdigest() == digest
actual += json.loads(data)
assert actual == expected, 'Generated variants differ from reviewed editorial sources'
assert len({entry['id'] for entry in actual}) == 12
for entry in actual:
task = entry['task']
base = next(x for x in json.loads((BANK / f'introductions.{task["skill"]}.json').read_text()) if x['id'] == entry['base_entry_id'])
assert task['target_language'] == task['instruction_language'] == 'en-US'
assert task['module_id'] == 'a1_introductions_v1' and task['instruction']
assert entry['equivalence'] in ('translation', 'adaptation')
assert (entry['canonical_task_id'] == entry['base_entry_id']) == (entry['equivalence'] == 'translation')
for field in ('audio_text', 'passage', 'reference_text'):
assert task.get(field) == base['task'].get(field), (entry['id'], field)
if task['kind'] == 'choice':
assert task['answer_key']['correct_option_id'] == base['task']['answer_key']['correct_option_id']
assert len({o['label'].casefold() for o in task['options']}) == 3
if entry['equivalence'] == 'translation':
assert [o['label'] for o in task['options'] if o['role'] == 'target'] == [o['label'] for o in base['task']['options']]
assert all(o['role'] in ('target', 'meaning') and o['language'] for o in task['options'])
if task['skill'] == 'listening':
assert not task.get('stimulus'), 'A private recording must never become a public stimulus'
assert sum(x['equivalence'] == 'adaptation' for x in actual) == 3
assert {x['task']['kind'] for x in actual} == {'choice', 'repeat_sentence', 'free_speech', 'short_text', 'free_text'}
assert all(sum(x['task']['skill'] == skill for x in actual) == 2 for skill in ('grammar', 'vocabulary', 'listening', 'reading', 'speaking', 'writing'))
print('PASS: 12 pilot variants; 9 equivalent translations, 3 separate adaptations; preserved keys/media; six learning areas and five renderers.')
if __name__ == '__main__':
build()
audit()