| """Validate an benchmark inputs and published per-example results.""" |
| import collections |
| import hashlib |
| import json |
| from pathlib import Path |
| import pyarrow.parquet as pq |
| from huggingface_hub import DatasetCard |
|
|
| ROOT=Path(__file__).resolve().parents[1] |
| VARIANTS=('original','reversed','serialization_control') |
| FORBIDDEN={'predicted_label','chosen_key','correct','confidence','probability_0','probability_1','response_id','evaluated'} |
|
|
|
|
| def validate(root=ROOT): |
| manifest=json.loads((root/'SHA256SUMS.json').read_text()) |
| actual={str(p.relative_to(root)) for p in root.rglob('*') |
| if p.is_file() and p.name!='SHA256SUMS.json' and '__pycache__' not in p.parts} |
| assert actual==set(manifest),'Unlisted or missing publication files' |
| for name,checksum in manifest.items(): |
| assert hashlib.sha256((root/name).read_bytes()).hexdigest()==checksum,name |
| card=DatasetCard.load(str(root/'README.md')) |
| assert {c['config_name'] for c in card.data.configs}=={'requests','case_index','model_results'} |
| configs={c['config_name']:c['data_files'] for c in card.data.configs} |
| assert configs['requests']==[dict(split=v,path=f'data/requests-{v}.parquet') for v in VARIANTS] |
| assert configs['case_index']==[dict(split='cases',path='data/case-index.parquet')] |
| expected={f'data/requests-{v}.parquet' for v in VARIANTS}|{'data/case-index.parquet'} |
| result_counts=dict(original=415827,reversed=415827,serialization_control=83183,repeat=246313,unsupported=352) |
| expected.update(f'data/model-results-{v}.parquet' for v in result_counts) |
| assert configs['model_results']==[dict(split=v,path=f'data/model-results-{v}.parquet') for v in result_counts] |
| for variant,count in result_counts.items(): |
| assert pq.ParquetFile(root/f'data/model-results-{variant}.parquet').metadata.num_rows==count |
| assert {name for name in manifest if name.startswith('data/')}==expected |
| tables={} |
| for variant in VARIANTS: |
| table=pq.read_table(root/'data'/f'requests-{variant}.parquet') |
| assert not FORBIDDEN.intersection(table.column_names) |
| assert table.num_rows==83183 |
| tables[variant]=table.to_pylist() |
| original={r['pair_id']:r for r in tables['original']} |
| assert len(original)==83183 |
| for i,a in enumerate(tables['original']): |
| b=tables['reversed'][i];c=tables['serialization_control'][i] |
| for row in [a,b,c]: |
| assert row['pair_id']==a['pair_id'] and row['state']==a['state'] |
| assert row['instruction']==a['instruction'] and row['gold_label']==a['gold_label'] |
| assert row['state_available'] and row['state'] |
| assert hashlib.sha256(row['state'].encode()).hexdigest()==row['state_sha256'] |
| assert next(o['semantic_label'] for o in row['options'] if o['key']==row['gold_key'])==row['gold_label'] |
| assert [o['key'] for o in a['options']]==[o['key'] for o in b['options']]==['A','B'] |
| assert [o['text'] for o in a['options']]==[o['text'] for o in b['options']][::-1] |
| assert a['options']==c['options'][::-1] |
| groups=collections.defaultdict(list) |
| for row in original.values():groups[row['state_sha256']].append(row) |
| index=pq.read_table(root/'data/case-index.parquet') |
| assert not FORBIDDEN.intersection(index.column_names) |
| rows=index.to_pylist();assert len(rows)==83183 and len({r['pair_id'] for r in rows})==83183 |
| for row in rows: |
| source=original[row['pair_id']];group=groups[source['state_sha256']] |
| assert row['content_group_id']==source['state_sha256'] |
| assert row['benchmark']==source['benchmark'] and row['duplicate_rows']==len(group) |
| assert row['duplicate_benchmarks']==sorted({r['benchmark'] for r in group}) |
| assert row['label_conflict']==(len({r['gold_label'] for r in group})>1) |
| assert row['rubric_conflict']==(len({r['instruction'] for r in group})>1) |
| text=(root/'README.md').read_text() |
| for section in ['## Results','### Correct answers: original options','### Wrong answers: original options', |
| '### Correct answers: reversed options','### Correct answers: order-only test', |
| '### Did changing the options change the answer?','### Results at a glance']: |
| assert section in text |
| for name in ['Original Jev','CLM','Laya','Open-Jev','Laya Multilingual']:assert name in text |
| return dict(passed=True,request_rows=249549,case_index_rows=83183, |
| model_result_dataset_rows=1161502,result_location='model_results and README.md', |
| full_text_harmbench_rows_per_variant=sum(r['benchmark'].startswith('HarmBench') for r in original.values()), |
| checksum_files=len(manifest)) |
|
|
|
|
| if __name__=='__main__':print(json.dumps(validate(),indent=2)) |
|
|
