Publish Aussie-1k 0.4: 1500 questions (500 reviewed additions), dataset card and provenance
27c633b verified | #!/usr/bin/env python3 | |
| """Append a frozen, sourced and reviewed 500-question supplement to unchanged v6.2.""" | |
| import hashlib | |
| import json | |
| import unicodedata | |
| from collections import Counter | |
| from pathlib import Path | |
| ROOT = Path(__file__).resolve().parents[1] | |
| BASE_SHA = 'aaceb71c48331b813049f35bdb103d630f4acd46caeea2db831108f962f17630' | |
| AUTHOR = 'Claude Opus 5.5 (Anthropic), via parallel authoring agents' | |
| DATE = '2026-10-14' | |
| SELECTION = ('Authored and reviewed before any model was evaluated on these additions; ' | |
| 'no performance-based item selection') | |
| def sha(data): | |
| return hashlib.sha256(data).hexdigest() | |
| def build(): | |
| base = ROOT / 'data/v6.2/test.jsonl' | |
| assert sha(base.read_bytes()) == BASE_SHA | |
| rows = [json.loads(s) for s in base.read_text().splitlines()] | |
| source = [json.loads(s) for s in (ROOT / 'source/v6.3-additions.jsonl').read_text().splitlines()] | |
| assert len(source) == 500 | |
| audit = [] | |
| for line, item in enumerate(source, 1): | |
| question = unicodedata.normalize('NFC', item['question']).strip() | |
| qid = 'aussieqa-' + sha(question.encode())[:20] | |
| rows.append({'id': qid, 'question': question, 'answers': [item['answer']]}) | |
| audit.append({'id': qid, 'source_version': '6.3-additions', 'source_line': line, | |
| 'topic': item['topic'], 'subtopic': item['subtopic'], | |
| 'difficulty_estimate': item['difficulty_estimate'], 'source_url': item['source_url'], | |
| 'author': AUTHOR, 'authored_date': DATE, 'answer_source_checked_date': DATE, | |
| 'review': item['review'], 'selection': SELECTION}) | |
| assert len(rows) == len({r['id'] for r in rows}) == len({r['question'].casefold() for r in rows}) == 1500 | |
| for path, data in [('data/v6.3/test.jsonl', rows), ('metadata/v6.3/additions.jsonl', audit)]: | |
| p = ROOT / path | |
| p.parent.mkdir(parents=True, exist_ok=True) | |
| p.write_text(''.join(json.dumps(r, ensure_ascii=False, separators=(',', ':')) + '\n' for r in data)) | |
| assert (ROOT / 'data/v6.3/test.jsonl').read_bytes().startswith(base.read_bytes()) | |
| outcomes = Counter(item['review']['outcome'] for item in source) | |
| manifest = {'dataset': 'Aussie-1k', 'version': '0.4', 'source_build_identifier': '6.3', 'release_date': DATE, | |
| 'base_version': '0.3', 'unchanged_base_questions': 1000, 'new_questions': 500, | |
| 'test_records': 1500, 'reference_answers': sum(len(r['answers']) for r in rows), | |
| 'answer_count_distribution': dict(Counter(len(r['answers']) for r in rows)), | |
| 'addition_topic_counts': dict(Counter(r['topic'] for r in source)), | |
| 'addition_subtopic_slices': dict(Counter(r['slice'] for r in source)), | |
| 'addition_difficulty_estimates': dict(sorted(Counter(r['difficulty_estimate'] for r in source).items())), | |
| 'review_outcomes': dict(outcomes), | |
| 'review_protocol': ('Claude authored each item and checked it against the cited source. MiniMax-M3 and ' | |
| 'GLM-5.3-Flash (SCX.ai) independently rated fit (1-5), reference correctness, ambiguity ' | |
| 'and time-stability. Items accepted by both judges with no flags passed review; every ' | |
| 'flagged item was adjudicated by Claude against sources and kept unchanged, revised and ' | |
| 're-reviewed by both judges, or dropped. Claude then read every remaining candidate and ' | |
| 'dropped overlapping, answer-leaking or weakly attested items.'), | |
| 'selection_rule': ('From the reviewed candidates, 45 items per authoring slice (50 for geography); within ' | |
| 'each slice, difficulty levels keep proportional shares and items are ranked by the ' | |
| 'judges\' summed fit scores. No evaluated-model answers were used.'), | |
| 'candidate_counts': {'drafted': 608, 'dropped_as_duplicates_or_unstable_before_review': 14, | |
| 'dropped_in_adjudication_or_manual_review': 12, 'reserve_not_selected': 82, | |
| 'selected': 500}, | |
| 'duplicate_review': ('Exact and case-insensitive uniqueness checked; semantic near-duplicates against ' | |
| 'v6.2 and within the supplement flagged with E5-Mistral-7B-Instruct embeddings ' | |
| '(cosine >= 0.82) and resolved by hand. No claim of exhaustive semantic deduplication.'), | |
| 'ordering': 'Unchanged v6.2 questions followed by 500 additions in frozen source order', | |
| 'files': {}} | |
| for name in ['data/v6.2/test.jsonl', 'source/v6.3-additions.jsonl', 'data/v6.3/test.jsonl', | |
| 'metadata/v6.3/additions.jsonl']: | |
| b = (ROOT / name).read_bytes() | |
| manifest['files'][name] = {'sha256': sha(b), 'bytes': len(b)} | |
| (ROOT / 'metadata/v6.3/manifest.json').write_text(json.dumps(manifest, indent=2) + '\n') | |
| print(json.dumps(manifest, indent=2)) | |
| if __name__ == '__main__': | |
| build() | |