| """Conservatively exclude training template variants flagged by the teacher audit.""" |
| import argparse |
| from collections import Counter |
| import json |
| from pathlib import Path |
|
|
|
|
| def main(): |
| p=argparse.ArgumentParser();p.add_argument('--data',required=True);p.add_argument('--audit',required=True) |
| args=p.parse_args();base=Path(args.data);bad=set();accepted=0;jobs=0 |
| for line in Path(args.audit).read_text().splitlines(): |
| row=json.loads(line);jobs+=1;accepted+=len(row['accepted']) |
| for record in row['rejected']: |
| |
| tail=record['id'][6:] |
| import re |
| match=re.fullmatch(r'(.+)_(\d+)_(en|noisy_en|hi|hinglish)',tail) |
| assert match,record['id'] |
| op,index,language=match.groups();bad.add((op,language,int(index))) |
| removed=Counter();kept=0;source=base/'train.jsonl';temp=base/'train.filtered.jsonl' |
| with source.open() as inp,temp.open('w') as out: |
| for line in inp: |
| row=json.loads(line);key=(row['operation'],row['language'],row.get('template_index')) |
| if key in bad and row['provenance'].startswith('programmatic semantics + Qwen3.8-27B-FP8 language template'): |
| removed['/'.join(map(str,key))]+=1 |
| else:out.write(line);kept+=1 |
| temp.replace(source) |
| report={'audit_jobs':jobs,'accepted_template_instances':accepted,'flagged_template_variants':len(bad), |
| 'excluded_training_rows':sum(removed.values()),'remaining_training_rows':kept,'exclusions':dict(removed), |
| 'policy':'Conservative exclusion after SQL round-trip disagreement; disagreement alone does not prove the teacher was correct. Validation, test, manual, direct verified paraphrases and separately authored contrasts remain unchanged.', |
| 'limitation':'Earlier curriculum stages trained on these rows before the audit; this filtering does not undo earlier exposure.'} |
| (base/'template-filter.json').write_text(json.dumps(report,indent=2));print(json.dumps(report)) |
|
|
|
|
| if __name__=='__main__':main() |
|
|
