| """Explore the released benchmark without generating or executing model code. |
| Install: python -m pip install pandas pyarrow |
| Run: python quickstart.py |
| For a local release: python quickstart.py --local-release /path/to/release |
| """ |
| from pathlib import Path |
| import argparse |
| import json |
| import pandas as pd |
|
|
| REPO = 'uwm-software-engineering/complexity-kink-research' |
| REVISION = '1d1f538301d98c6e775e2ef12d9630b3b6bfd61d' |
|
|
|
|
| def load_tables(local_release=None): |
| """Select just the columns used by this example, from a pinned public revision.""" |
| wanted = { |
| 'prompts': ['prompt_id', 'ens_composite', 'construction_frame', 'task_type', |
| 'mean_pass_rate', 'mean_harness_pass_rate'], |
| 'generations': ['prompt_id', 'model_key', 'pass_rate', 'pass_rate_reviewed', |
| 'harness_pass_rate'], |
| } |
| tables = {} |
| for config, columns in wanted.items(): |
| suffix = f'{config}/test-00000-of-00001.parquet' |
| source = str(Path(local_release) / suffix) if local_release else ( |
| f'https://huggingface.co/datasets/{REPO}/resolve/{REVISION}/{suffix}') |
| tables[config] = pd.read_parquet(source, columns=columns) |
| return tables['prompts'], tables['generations'] |
|
|
|
|
| def analyze(prompts, generations): |
| assert prompts['prompt_id'].is_unique |
| assert not generations.duplicated(['prompt_id', 'model_key']).any() |
| assert len(prompts) == 5000 and len(generations) == 105000 |
| assert generations['model_key'].nunique() == 21 |
| assert generations.groupby('prompt_id')['model_key'].nunique().eq(21).all() |
| joined = generations.merge(prompts[['prompt_id', 'construction_frame']], |
| on='prompt_id', how='left', validate='many_to_one') |
| assert joined['construction_frame'].notna().all() |
| |
| |
| outcomes = generations.groupby('model_key', sort=True).agg( |
| generations=('prompt_id', 'size'), |
| audited_primary=('pass_rate', 'mean'), |
| reviewed_version=('pass_rate_reviewed', 'mean'), |
| raw_harness=('harness_pass_rate', 'mean')) |
| outcomes['primary_minus_harness_pp'] = ( |
| outcomes['audited_primary'] - outcomes['raw_harness']) * 100 |
| p = prompts.copy() |
| |
| p['above_reported_threshold'] = p['ens_composite'] > 14.0 |
| frame = p.groupby(['construction_frame', 'above_reported_threshold']).agg( |
| prompts=('prompt_id', 'size'), |
| mean_index=('ens_composite', 'mean'), |
| audited_primary=('mean_pass_rate', 'mean'), |
| raw_harness=('mean_harness_pass_rate', 'mean')) |
| inventory = {'prompts': len(prompts), 'generations': len(generations), |
| 'evaluated_models': int(generations['model_key'].nunique()), |
| 'mean_audited_primary': float(generations['pass_rate'].mean()), |
| 'mean_reviewed_version': float(generations['pass_rate_reviewed'].mean()), |
| 'mean_raw_harness': float(generations['harness_pass_rate'].mean())} |
| return inventory, outcomes, frame |
|
|
|
|
| def main(): |
| parser = argparse.ArgumentParser(description=__doc__) |
| parser.add_argument('--local-release', type=Path) |
| parser.add_argument('--output', type=Path) |
| args = parser.parse_args() |
| inventory, outcomes, frame = analyze(*load_tables(args.local_release)) |
| print('BENCHMARK INVENTORY\n' + json.dumps(inventory, indent=2)) |
| print('\nOUTCOME DEFINITIONS BY MODEL (alphabetical, not a current leaderboard)') |
| print(outcomes.round(4).to_string()) |
| print('\nCONSTRUCTION FRAME AND THE REPORTED POOLED THRESHOLD (descriptive)') |
| print(frame.round(4).to_string()) |
| print('\nInterpretation: differences can reflect test quality, task mix, model settings,') |
| print('and generation runs. They do not establish causal effects, a universal failure') |
| print('threshold, or reliability on an unseen customer task. No generated code was run.') |
| if args.output: |
| result = {'inventory': inventory, |
| 'model_outcome_definitions': outcomes.reset_index().to_dict('records'), |
| 'frame_summary': frame.reset_index().to_dict('records'), |
| 'public_dataset_revision': REVISION, |
| 'loaded_from': str(args.local_release) if args.local_release else 'public pinned release'} |
| args.output.write_text(json.dumps(result, indent=2) + '\n') |
|
|
|
|
| if __name__ == '__main__': |
| main() |
|
|
