File size: 3,255 Bytes
48f76bc | 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 63 64 65 66 67 | """Audit omissions including unlabeled units, and propose a supported allowlist.
Usage: guide_audit.py TRAIN_JSONL PREDICTIONS_JSONL OUT_JSON
Uses the installed B2 thresholds (.38 probability, .5 confidence). A unit must
have both label classes in training and validation, at least five examples of
each on validation, and at least five omissions at 97% precision with 98% needed-guide retention
to qualify.
Labels are observational proxies; this does not establish instruction usefulness.
"""
import collections
import json
import sys
from pathlib import Path
def audit(train, predictions, threshold=.38, confidence=.5):
training = collections.defaultdict(collections.Counter)
for row in train:
for name, value in row['labels']['guides'].items():
if value is not None:
training[name]['positive' if value else 'negative'] += 1
stats = collections.defaultdict(collections.Counter)
for row in predictions:
for name, value in row['answers'].get('guides', {}).get('probabilities', {}).items():
needed = row['labels']['guides'].get(name)
omitted = value < threshold and max(value, 1 - value) >= confidence
stats[name]['scored'] += 1
stats[name]['omitted'] += omitted
if needed is None:
stats[name]['unknown'] += 1
stats[name]['omitted_unknown'] += omitted
else:
stats[name]['positive' if needed else 'negative'] += 1
stats[name]['omitted_needed' if needed else 'omitted_unneeded'] += omitted
allowed = []
for name, s in stats.items():
known_omissions = s['omitted_needed'] + s['omitted_unneeded']
if (training[name]['positive'] > 0 and training[name]['negative'] > 0
and s['positive'] >= 5 and s['negative'] >= 5 and not s['omitted_unknown']
and known_omissions >= 5 and s['omitted_unneeded'] / known_omissions >= .97
and 1 - s['omitted_needed'] / s['positive'] >= .98):
allowed.append(name)
return {'omittable': sorted(allowed), 'omit_below': threshold, 'confidence_floor': confidence,
'training': {k: dict(v) for k, v in sorted(training.items())},
'validation': {k: dict(v) for k, v in sorted(stats.items())},
'limitation': 'Observed tool use supplies guide labels; unlabeled guidance cannot be certified for omission.'}
def main():
train, predictions, output = sys.argv[1:]
def read(path):
return [json.loads(s) for s in Path(path).read_text().splitlines() if s.strip()]
training, scored = read(train), read(predictions)
candidates = [audit(training, scored, step / 100) for step in range(1, 39)]
result = max(candidates, key=lambda r: sum(r['validation'][n]['omitted'] for n in r['omittable']))
if not result['omittable']:
result['omit_below'] = 0
result['original_threshold_audit'] = audit(training, scored)
result['selection_rule'] = 'Maximize supported omissions on validation; per-unit precision >= .97 and needed retention >= .98.'
Path(output).write_text(json.dumps(result, indent=2) + '\n')
print(json.dumps(result, indent=2))
if __name__ == '__main__':
main()
|