File size: 3,181 Bytes
ad506db | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 | """Read-only aggregation of explicitly selected existing alias decisions.
Never calls an LLM, changes an answer, infers aliases, or silently chooses a judge.
"""
import argparse,collections,json
from pathlib import Path
def summarize(rows, field, error_policy):
groups=collections.defaultdict(lambda:dict(rows=0,scored=0,correct=0,errors=0,unscored=0))
seen=set()
for r in rows:
qid=r.get('qa_id');condition=r.get('condition',r.get('setting'))
if not qid or not condition:raise ValueError('Every answer must have qa_id and condition/setting')
key=(qid,condition)
if key in seen:raise ValueError(f'Duplicate answer key: {key}; resolve explicitly before scoring')
seen.add(key)
error=bool(r.get('error') or r.get('api_error'))
value=r.get(field)
for group in [(condition,'Overall'),(condition,str(r.get('subtype') or 'unknown'))]:
g=groups[group];g['rows']+=1;g['errors']+=int(error)
if error and error_policy=='exclude':continue
if error and error_policy=='count_as_incorrect':value=False
if not isinstance(value,bool):g['unscored']+=1;continue
g['scored']+=1;g['correct']+=int(value)
result=[]
for (c,s),g in sorted(groups.items()):
result.append(dict(condition=c,subtype=s,**g,accuracy_pct=100*g['correct']/g['scored'] if g['scored'] else None))
return dict(correct_field=field,error_policy=error_policy,unique_answer_keys=len(seen),groups=result)
def main():
p=argparse.ArgumentParser(description=__doc__)
p.add_argument('--answers',required=True,type=Path)
p.add_argument('--correct-field',required=True)
p.add_argument('--error-policy',required=True,choices=['exclude','count_as_incorrect','stored'])
p.add_argument('--qa',type=Path,help='Optional exact evaluation QA subset; joins source_medium_qa_id or qa_id')
p.add_argument('--out',required=True,type=Path)
a=p.parse_args(); qa={}
if a.qa:
with a.qa.open(encoding='utf-8-sig') as f:
for line in f:
if not line.strip():continue
r=json.loads(line);qid=r.get('source_medium_qa_id') or r['qa_id']
if qid in qa:raise ValueError(f'Duplicate QA subset ID: {qid}')
qa[qid]=r
if not qa:raise ValueError('Explicit QA subset is empty; refusing full-set fallback')
def records():
with a.answers.open(encoding='utf-8-sig') as f:
for line in f:
if not line.strip():continue
r=json.loads(line)
if qa:
qid=r.get('source_medium_qa_id') or r['qa_id']
if qid not in qa:continue
r=dict(r,subtype=qa[qid].get('subtype',r.get('subtype')))
yield r
result=summarize(records(),a.correct_field,a.error_policy)
result['qa_subset_size']=len(qa) if a.qa else None
a.out.parent.mkdir(parents=True,exist_ok=True)
a.out.write_text(json.dumps(result,ensure_ascii=False,indent=2),encoding='utf-8')
print(json.dumps({'output':str(a.out),'keys':result['unique_answer_keys']}))
if __name__=='__main__':main()
|