scripts/evaluate.py

"""Score explicit option-letter predictions against the frozen SUM release.

Usage: python scripts/evaluate.py --annotations dataset.json --predictions predictions.jsonl
Each prediction is {"id":"000001", "prediction":"A"}; absent/invalid answers count as incorrect.
"""
import argparse
import collections
import json
from pathlib import Path

p=argparse.ArgumentParser()
p.add_argument('--annotations',type=Path,required=True)
p.add_argument('--predictions',type=Path,required=True)
a=p.parse_args()
rows=json.loads(a.annotations.read_text())
predictions={}
for line in a.predictions.read_text().splitlines():
    if not line.strip():continue
    r=json.loads(line)
    if r['id'] in predictions:raise ValueError('Duplicate prediction ID: '+r['id'])
    predictions[r['id']]=str(r['prediction']).strip().upper()
known={r['id'] for r in rows}
if set(predictions)-known:raise ValueError('Unknown prediction IDs')
groups=collections.defaultdict(lambda:dict(correct=0,total=0,missing=0,invalid=0))
for r in rows:
    answer=predictions.get(r['id'])
    valid={c['label'] for c in r['choices']}
    for key in ['overall','source/'+r['source_dataset'],'task/'+r['task_type']]:
        s=groups[key];s['total']+=1
        s['correct']+=int(answer==r['answer'])
        s['missing']+=int(answer is None)
        s['invalid']+=int(answer is not None and answer not in valid)
for s in groups.values():s['accuracy']=s['correct']/s['total']
print(json.dumps(groups,indent=2))