File size: 2,550 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 | """Minimal exact-answer demo, not the paper's complete biomedical alias judge."""
import argparse, collections, gzip, json, unicodedata
from pathlib import Path
def records(path):
opener=gzip.open if str(path).endswith('.gz') else open
with opener(path,'rt',encoding='utf-8') as f:
for line in f:
if line.strip():yield json.loads(line)
def norm(value):
return ' '.join(unicodedata.normalize('NFKC',str(value)).casefold().split())
def evaluate(qa,predictions,aliases=None):
gold={}
for row in qa:
qid=row['qa_id']
if qid in gold:raise ValueError('Duplicate QA ID')
gold[qid]=row
counts=collections.Counter(total_qa=len(gold));seen=set()
for row in predictions:
qid=row['qa_id']
if qid in seen:raise ValueError('Duplicate prediction ID; use one condition per file')
if qid not in gold:raise ValueError('Unknown prediction QA ID')
seen.add(qid)
if row.get('error') or row.get('api_error'):
counts['api_errors']+=1;continue
prediction=row.get('answer')
if isinstance(prediction,list) and len(prediction)==1:prediction=prediction[0]
if not isinstance(prediction,str) or not prediction.strip():
counts['invalid_predictions']+=1;continue
target=gold[qid].get('short_answer',gold[qid].get('answer'))
if isinstance(target,list) and len(target)==1:target=target[0]
if not isinstance(target,str):raise ValueError('Demo requires a single string answer')
accepted=[target]+(aliases or {}).get(qid,[])
counts['scored']+=1
counts['correct']+=int(norm(prediction) in {norm(x) for x in accepted})
counts['missing_predictions']=len(gold)-len(seen)
return dict(counts,accuracy_pct=100*counts['correct']/counts['scored'] if counts['scored'] else None,
scoring='NFKC + casefold + whitespace exact match; optional explicit QA aliases',
error_policy='exclude API errors and invalid/missing predictions; report counts')
def main():
p=argparse.ArgumentParser(description=__doc__)
p.add_argument('--qa',required=True,type=Path)
p.add_argument('--predictions',required=True,type=Path)
p.add_argument('--aliases',type=Path,help='JSON object: QA ID -> list of accepted answer strings')
a=p.parse_args()
aliases=json.loads(a.aliases.read_text(encoding='utf-8')) if a.aliases else None
print(json.dumps(evaluate(records(a.qa),records(a.predictions),aliases),indent=2))
if __name__=='__main__':main()
|