Download appendix/code/score_distillation.py from BradJu/attribute-ownership-negative-experiments: direct link, hf CLI and curl.
- Browser
- Download file 2.13 kB
-
https://huggingface.co/datasets/BradJu/attribute-ownership-negative-experiments/resolve/main/appendix/code/score_distillation.py
- Command line
-
hf download hf://datasets/BradJu/attribute-ownership-negative-experiments/appendix/code/score_distillation.py
-
curl -L -o score_distillation.py https://huggingface.co/datasets/BradJu/attribute-ownership-negative-experiments/resolve/main/appendix/code/score_distillation.py
2.13 kB
| import argparse,json | |
| from pathlib import Path | |
| from metrics import summarize,bootstrap | |
| from environment import Shop | |
| R=Path(__file__).resolve().parent | |
| p=argparse.ArgumentParser();p.add_argument('outputs',type=Path);a=p.parse_args() | |
| rows=[r for r in json.loads((R.parent/'data/tasks.json').read_text()) if r['split']=='holdout'] | |
| results={c:json.loads((a.outputs/(c+'.json')).read_text())['results'] for c in ['base_agent','distilled_agent','gate_controller']} | |
| summaries={c:summarize(rows,r) for c,r in results.items()};interval=bootstrap(rows,results['base_agent'],results['distilled_agent']);b=summaries['base_agent'];d=summaries['distilled_agent'] | |
| criteria={'gain_15pp_and_positive_ci':(d['success']-b['success'])/72>=.15 and interval[0]>0,'wrong_filter_not_worse':d['wrong_filter_count']<=b['wrong_filter_count'],'explicit_recall_60pct':d['explicit_retained']/d['explicit_total']>=.6,'parse_failures_at_most_20pct':d['parse_failures']/72<=.2} | |
| controls={} | |
| for name in ['constant_no','four_phrase']: | |
| predictions=[];matches=0 | |
| for row in rows: | |
| text=row['input']['text'].lower();gate=next((g for phrase,g in [('must have','YES'),('do not require','NO'),('not decided','UNSURE'),('compatible with','FIT')] if phrase in text),None) if name=='four_phrase' else 'NO' | |
| matches+=gate is not None | |
| if gate is None:predictions.append({'id':row['id'],'success':False,'error':'NO_MATCH','calls':[],'steps':0});continue | |
| proxy={**row,'status':{'YES':'explicit','NO':'not_stated','FIT':'compatibility_only','UNSURE':'ambiguous'}[gate]} | |
| env=Shop(proxy);env.teacher();env.row=row | |
| predictions.append({'id':row['id'],'success':env.success(),'error':None,'calls':env.calls,'steps':len(env.calls)}) | |
| controls[name]={'matched':matches,'metrics':summarize(rows,predictions)} | |
| report={'controls':controls,'summaries':summaries,'paired_family_bootstrap_95':interval,'adoption_criteria':criteria,'adopt':all(criteria.values()),'limitations':'Rule-teacher SFT; synthetic exposed holdout; simulated catalog; model-assisted labels.'} | |
| (a.outputs/'local-results.json').write_text(json.dumps(report,indent=2)+'\n');print(json.dumps(report,indent=2)) | |