| """Frozen matched training contrast and one ensemble replacement; no Test.""" |
| import csv |
| import json |
| from pathlib import Path |
| import sys |
| ROOT=Path('/marimo/storage/road');sys.path.insert(0,str(ROOT)) |
| from visual_preference import policy,digest,write_json,RUN |
| from lighton_completed_gate_oct2 import BUNDLE,ORDER,table,compare,rover,drop_runaways |
|
|
| def main(): |
| freeze,split,source,fit_ids=policy();ids=split['val_ids'] |
| run=json.loads((RUN/'run_config.json').read_text()) |
| assert run['fit_ids']==fit_ids and run['val_ids']==ids and run['validation_contaminated'] is False |
| assert run['training_source_sha256']==digest(ROOT/'visual_preference.py') |
| marker=json.loads((RUN/'final/SAVE_COMPLETE').read_text()) |
| assert marker['step']==freeze['configuration']['steps'] |
| assert marker['weight_sha256']==digest(RUN/'final/adapter_model.safetensors') |
| assert marker['manifest_sha256']==digest(RUN/'final/run_config.json') |
| for kind in ['initial','final']: |
| sidecar=json.loads((ROOT/f'preds_val_visual_preference_{kind}_oct2.manifest.json').read_text()) |
| assert sidecar['ids']==ids and sidecar['freeze_sha256']==digest(ROOT/'visual_preference_oct2.freeze.json') |
| assert sidecar['source_sha256']==digest(ROOT/'visual_preference.py') |
| expected=freeze['initialization']['weight_sha256'] if kind=='initial' else marker['weight_sha256'] |
| assert sidecar['adapter_sha256']==expected |
| for i,sha in sidecar['image_sha256'].items():assert digest(ROOT/'images'/(i+'.jpg'))==sha |
| initial=ROOT/'preds_val_visual_preference_initial_oct2.csv' |
| final=ROOT/'preds_val_visual_preference_final_oct2.csv' |
| matched=compare(initial,final,BUNDLE/'Train.csv',BUNDLE/'split_manifest.json',10000,20261002) |
| tables={name:table(BUNDLE/('m_'+name+'.csv'),ids) for name in ORDER} |
| tables['preference']=table(final,ids) |
| scratch=ROOT/'visual_preference_gate_oct2';scratch.mkdir(exist_ok=True) |
| for name,members in [('control',ORDER),('replace_r64',ORDER[:1]+['preference']+ORDER[2:])]: |
| with (scratch/(name+'.csv')).open('w',newline='') as f: |
| w=csv.writer(f);w.writerow(['ID','raw']) |
| for i in ids:w.writerow([i,rover(drop_runaways([tables[m][i] for m in members if tables[m][i]]))]) |
| report=compare(scratch/'control.csv',scratch/'replace_r64.csv',BUNDLE/'Train.csv',BUNDLE/'split_manifest.json',10000,20261002) |
| assert abs(report['sqrt_weighted_raw']['baseline']-.9318177934195226)<1e-12 |
| def passed(r): |
| w=r['sqrt_weighted_raw'];return w['delta']>=.001 and w['paired_delta_ci95'][0]>0 and r['arithmetic_raw']['delta']>0 |
| report.update(matched_training=matched,decision='eligible_for_artifact_review' if passed(report) and passed(matched) else 'do_not_promote', |
| final_model_sha256=marker['weight_sha256'],freeze_sha256=digest(ROOT/'visual_preference_oct2.freeze.json'), |
| eligible_preference_pairs=run['eligible_preference_pairs'],test_queued=False,submission_attempted=False, |
| limitations='Repeated holdout; matched training and one fixed ensemble arm. Production artifact gaps still require review.') |
| write_json(ROOT/'visual_preference_completed_gate_oct2.json',report);print(json.dumps(report),flush=True) |
|
|
| if __name__=='__main__':main() |
|
|
