File size: 3,262 Bytes
08a9326 | 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 | """Descriptive pilot summaries exclusively from authoritative extension .eval logs."""
from pathlib import Path
import json
import pandas as pd
from inspect_ai.log import read_eval_log
from encoding_tax.common import ROOT,digest,write_json,now
from encoding_tax.scoring import auroc
def main():
root=ROOT/'results/e6-exploratory-v2/pilot/inspect';out=ROOT/'reports/e6-exploratory-v2/pilot';out.mkdir(parents=True,exist_ok=True)
observations={};hashes={}
for path in sorted(root.glob('*.eval')):
log=read_eval_log(path);hashes[str(path.relative_to(ROOT))]=digest(path)
if not log.eval.metadata.get('exploratory') or log.eval.task!='E6X2_pilot':raise ValueError('Mixed study logs')
for sample in log.samples or []:
m=sample.metadata;key=(m['reader'],sample.id)
row={k:m[k] for k in ['reader','mode','policy','representation','rounds','budget','label','cluster_id','leaves']}
row.update(sample_id=sample.id,error=bool(sample.error),raw=sample.output.completion if sample.output else '',scratchpad=m.get('scratchpad'),reasoning_end=m.get('reasoning_end'),**m.get('outcome',{}))
requests=m.get('requests',[])
row['prompt_hashes']=[r.get('prompt_sha256') for r in requests]
for field in ['completion_tokens','prompt_tokens','cost']:
values=[r.get(field) for r in requests];row[field]=sum(values) if len(values)==2 and all(x is not None for x in values) else None
if key in observations and not observations[key]['error']:
old=observations[key]
if not row['error'] and (old['raw']!=row['raw'] or old['prompt_hashes']!=row['prompt_hashes']):raise ValueError('Conflicting completed duplicate')
continue
observations[key]=row
frame=pd.DataFrame(observations.values());frame.to_json(out/'observations.jsonl',orient='records',lines=True)
rows=[]
keys=['reader','mode','policy','representation','rounds','budget']
for k,g in frame.groupby(keys,dropna=False):
good=g[~g.error];row=dict(zip(keys,k));row.update(n=len(g),operational_errors=int(g.error.sum()),completed=len(good),problems=g.cluster_id.nunique(),completion_tokens=good.completion_tokens.mean(),input_tokens=good.prompt_tokens.mean(),cost=good.cost.sum(min_count=1),validity=good.valid.mean(),scratchpad_cap_rate=good.reasoning_end.eq('length').mean())
if row['mode']=='monitor':
row.update(auroc=auroc(good.label,good.probability),recall=good[good.label==1].alert.mean(),fpr=good[good.label==0].alert.mean(),log_loss=good.log_loss.mean())
valid=good[good.valid==True];row['valid_only_auroc']=auroc(valid.label,valid.probability)
else:
for metric in ['decoded_exact','equation_correct','first_error_correct','all_correct']:row[metric]=good[metric].mean()
rows.append(row)
pd.DataFrame(rows).to_csv(out/'conditions.csv',index=False)
write_json(out/'provenance.json',{'created_at':now(),'exploratory':True,'descriptive_pilot_only':True,'expected_samples':9120,'observed_samples':len(frame),'operational_errors':int(frame.error.sum()),'source_log_hashes':hashes})
print('Pilot analysis:',out)
if __name__=='__main__':main()
|