| """Rerun one complete judge and compare every held-out rating to the archive.""" |
| import argparse |
| import hashlib |
| import json |
| from pathlib import Path |
|
|
| import numpy as np |
| import pandas as pd |
|
|
| from softelo_release.data import load_annotations |
| from softelo_release.reproduce import run_point_estimates |
|
|
| p=argparse.ArgumentParser(description=__doc__) |
| p.add_argument('--data',type=Path,required=True) |
| p.add_argument('--manifest-sha256',required=True) |
| p.add_argument('--judge',default='GPT-OSS-120B') |
| p.add_argument('--slug',default='gpt_oss_120b') |
| a=p.parse_args() |
| manifest=a.data/'manifest.json' |
| if hashlib.sha256(manifest.read_bytes()).hexdigest()!=a.manifest_sha256: |
| raise ValueError('Dataset manifest changed after verification recipe was recorded') |
| md=json.loads(manifest.read_text()) |
| spec=next(r for r in md['annotations'] if r['corpus']=='lmarena100k' and r['judge']==a.judge) |
| if hashlib.sha256((a.data/spec['file']).read_bytes()).hexdigest()!=spec['sha256']: |
| raise ValueError('Annotation checksum differs') |
| print(json.dumps({'judge':a.judge,'corpus':'lmarena100k','protocol':'main91_strict_loo', |
| 'manifest_sha256':a.manifest_sha256,'annotation_sha256':spec['sha256']}),flush=True) |
| frame=load_annotations(a.data,corpus='lmarena100k',judge=a.judge) |
| print(f'Loaded {len(frame)} valid annotation rows',flush=True) |
| summary, detail=run_point_estimates(frame) |
| expected=pd.read_csv(a.data/'results/main91'/a.slug/'heldout_models_decisive.csv') |
| expected=expected[expected.method.isin(['hard','temperature_mle'])] |
| keys=['method','model'];left=detail.set_index(keys).sort_index();right=expected.set_index(keys).sort_index() |
| if not left.index.equals(right.index):raise ValueError('Held-out model sets differ') |
| metrics={} |
| for column in ['judge_bt_elo','human_bt_elo','temperature']: |
| x=left[column].to_numpy(float);y=right[column].to_numpy(float) |
| metrics[column+'_max_abs_difference']=float(np.nanmax(np.abs(x-y))) |
| if not np.allclose(x,y,rtol=0,atol=1e-5,equal_nan=True): |
| print(json.dumps(metrics),flush=True) |
| raise ValueError(f'Numerical reproduction differs for {column}') |
| print(summary.to_string(index=False),flush=True) |
| print(json.dumps({'passed':True,'heldout_rows':len(detail),**metrics}),flush=True) |
| Path('release-check-output').mkdir(exist_ok=True) |
| summary.to_csv('release-check-output/summary.csv',index=False) |
| detail.to_csv('release-check-output/heldout_models.csv',index=False) |
| Path('release-check-output/validation.json').write_text(json.dumps({'passed':True,**metrics},indent=2)+'\n') |
|
|
