empire-eval / scripts /package.py
Lilambd's picture
Publish Lilambd Empire Eval
7f9a652 verified
Raw History Blame Contribute Delete
5.13 kB
"""Split built questions into the public package and the private held-out answers.
usage: python package.py --built OUT_DIR --pkg PKG_DIR --private PRIVATE_DIR
PKG_DIR/data/dev/<set>.jsonl public: question + answer + evidence + recipe (70%)
PKG_DIR/data/test/<set>.jsonl public: question only (30%); answers are not published
PRIVATE_DIR/test_answers/<set>.jsonl the held-out answers (never upload)
The split is a fixed hash of the question id, so it is stable across rebuilds.
"""
import argparse, collections, glob, hashlib, json, os, shutil
PUBLIC_TEST_FIELDS = ['id', 'dataset', 'revision', 'task', 'difficulty', 'skills', 'question', 'answer_type', 'unit', 'answer_format', 'tolerance']
def _tol(t):
f = lambda v: None if v is None else float(v)
return {'abs': f(t.get('abs')), 'rel': f(t.get('rel'))}
def flat(it):
"""Arrow/HF-friendly row: mixed-type fields become JSON strings."""
o = dict(it)
if 'answer' in o:
v = o['answer']
o['answer_json'] = json.dumps(v, ensure_ascii=False)
o['answer'] = None if v is None else (', '.join(map(str, v)) if isinstance(v, list) else str(v))
if 'recipe' in o:
o['recipe'] = None if o['recipe'] is None else json.dumps(o['recipe'], ensure_ascii=False)
if 'evidence_scope' in o:
o['evidence_scope'] = json.dumps(o['evidence_scope'], ensure_ascii=False)
if 'evidence' in o:
o['evidence'] = [{**e, 'key': json.dumps(e.get('key') or {}, ensure_ascii=False)} for e in o['evidence']]
if 'tolerance' in o:
o['tolerance'] = _tol(o['tolerance'])
for k in ('answer', 'recipe', 'evidence_scope', 'unit', 'answer_format', 'note', 'verified_by', 'source_qid'):
if o.get(k) is None:
o[k] = '' # empty string, never null: keeps one Arrow schema across files
o['evidence'] = [{'row_id': str(e.get('row_id') or ''), 'file': e.get('file') or '', 'row': int(e['row']) if e.get('row') is not None else -1, 'key': e['key']} for e in o.get('evidence', [])]
o['tolerance'] = o.get('tolerance') or {'abs': None, 'rel': None}
return o
def unflat(o):
it = dict(o)
if 'answer_json' in it:
it['answer'] = json.loads(it.pop('answer_json'))
for k in ('recipe', 'evidence_scope'):
if isinstance(it.get(k), str):
it[k] = json.loads(it[k]) if it[k] else None
for k in ('unit', 'answer_format', 'note', 'verified_by', 'source_qid'):
if it.get(k) == '':
it[k] = None
for e in it.get('evidence', []) or []:
if isinstance(e.get('key'), str):
e['key'] = json.loads(e['key'])
if isinstance(it.get('tolerance'), dict):
it['tolerance'] = {k: v for k, v in it['tolerance'].items() if v is not None}
return it
def is_test(qid):
return int(hashlib.sha256(qid.encode()).hexdigest()[:8], 16) % 10 < 3
def main():
ap = argparse.ArgumentParser()
ap.add_argument('--built', required=True)
ap.add_argument('--pkg', required=True)
ap.add_argument('--private', required=True)
a = ap.parse_args()
for d in ('data/dev', 'data/test'):
os.makedirs(os.path.join(a.pkg, d), exist_ok=True)
os.makedirs(os.path.join(a.private, 'test_answers'), exist_ok=True)
stats = collections.OrderedDict()
for p in sorted(glob.glob(os.path.join(a.built, 'data', '*.jsonl'))):
name = os.path.basename(p)[:-6]
dev, test, ans = [], [], []
for line in open(p, encoding='utf-8'):
it = json.loads(line)
if is_test(it['id']):
t = {k: (it.get(k) if it.get(k) is not None else '') for k in PUBLIC_TEST_FIELDS}
t['tolerance'] = {'abs': None, 'rel': None}
if it.get('tolerance'):
t['tolerance'] = _tol(it['tolerance'])
test.append(t)
ans.append(it)
else:
dev.append(flat(it))
for rows, path in ((dev, os.path.join(a.pkg, 'data/dev', name + '.jsonl')), (test, os.path.join(a.pkg, 'data/test', name + '.jsonl')),
(ans, os.path.join(a.private, 'test_answers', name + '.jsonl'))):
with open(path, 'w', encoding='utf-8') as f:
for r in rows:
f.write(json.dumps(r, ensure_ascii=False) + '\n')
stats[name] = {'dev': len(dev), 'test': len(test)}
man = json.load(open(os.path.join(a.built, 'manifest.json')))
old_p = os.path.join(a.pkg, 'manifest.json')
if os.path.exists(old_p): # adding sets to an existing package keeps the earlier ones
old = json.load(open(old_p))
old['datasets'].update(man['datasets'])
man = old
for k, v in stats.items():
man['datasets'].setdefault(k, {})['split'] = v
json.dump(man, open(os.path.join(a.pkg, 'manifest.json'), 'w'), indent=1, ensure_ascii=False)
tot = {s: sum(v[s] for v in stats.values()) for s in ('dev', 'test')}
print(json.dumps(tot), len(stats), 'sets')
if __name__ == '__main__':
main()