mlboydaisuke's picture
Core AI decision models on the 231 public JevBench items (Mac, 2026-09-24)
b7a859c
Raw History Blame Contribute Delete
1.35 kB
#!/usr/bin/env python3
"""Deterministic 30-row subset of the public 231 for a solo re-check of contended runs:
easy 10 (every 5th by id order), original 10 (4 choice, 4 noul, 2 score, spread by id), hard 10 by state length
(5 around the median, 3 at the 75–90th percentile, 2 longest). Records are copied unchanged."""
import json, os, sys
root = os.path.expanduser('~/code/jevbench/datasets/public')
def load(t): return [json.loads(l) for l in open(f'{root}/{t}.jsonl') if l.strip()]
out = []
easy = sorted(load('easy'), key=lambda r: r['id']); out += easy[::5][:10]
orig = sorted(load('original'), key=lambda r: r['id'])
for ty, k in (('choice', 4), ('noul', 4), ('score', 2)):
rows = [r for r in orig if r['question']['type'] == ty]; step = max(1, len(rows) // k); out += rows[::step][:k]
hard = sorted(load('hard'), key=lambda r: len(r['state'])); n = len(hard)
idx = [n // 2 - 2, n // 2 - 1, n // 2, n // 2 + 1, n // 2 + 2, int(n * .75), int(n * .82), int(n * .9), n - 2, n - 1]
out += [hard[i] for i in idx]
seen = set(); dst = sys.argv[1] if len(sys.argv) > 1 else 'recheck30.jsonl'
with open(dst, 'w') as fh:
for r in out:
if r['id'] in seen: continue
seen.add(r['id']); fh.write(json.dumps(r, ensure_ascii=False) + '\n')
print(dst, len(seen), 'rows;', 'hard state chars:', [len(hard[i]['state']) for i in idx])