import csv from dataclasses import dataclass import html from pathlib import Path import runpy import time from refined_v4.study import digest, read, write from .data import ROOT def summarize(case: dict): folder = Path(case['output']) points = [read(path) for path in (folder / 'eval').glob('*.json') if not path.name.startswith('test-')] budget = case.get('sft_tokens', 2097152) sft = [v for v in points if v['phase'] == 'sft' and v['counter'] == budget] if not sft and (folder / 'reused-sft.json').exists(): local = folder / 'eval/rl-0.json' source = Path(read(folder / 'reused-sft.json')['source_evaluation']) sft = [read(local if local.exists() else source)] rl = sorted([v for v in points if v['phase'] == 'rl'], key=lambda v: v['counter']) tracker = read(folder / 'wandb.json') if (folder / 'wandb.json').exists() else {} return dict(model=case['model'], task=case['task'], sft_tokens=budget, recipe=case.get('recipe', 'conditioned-relative'), sft=sft[0] if sft else None, rl=rl[-1] if rl else None, stopped=read(folder / 'stopped.json') if (folder / 'stopped.json').exists() else None, complete=(folder / 'complete.json').exists(), wandb=tracker.get('url', '')) def saved_test_comparisons(root: Path, plan: dict) -> str: from .verify import paired table = [] for case in plan['cases']: folder = Path(case['output']) if not (folder / 'complete.json').exists(): continue output = ROOT / 'verification' / f'{root.name}-{case["model"]}-{case["task"]}-native' if not (output / 'result.json').exists(): before = read(folder / 'eval' / f'test-sft-{plan["sft_tokens"]}.json') after = read(folder / 'eval' / f'test-rl-{plan["rl_rollouts"]}.json') write(output / 'protocol.json', dict(run=str(folder), selection='All prespecified sizes and both GSM variants; reuse exact terminal policies and their original held-out evaluations. No additional inference.', queries=before['metric']['queries'], samples=plan['eval_samples'], sft_sha256=digest(folder / 'sft.pt'), rl_sha256=digest(folder / 'rl.pt'), panel_sampling='Uniform hints and equal eligible difficulty bands; original frozen test panel.', panel=[dict(source_id=q['source_id'], group=q['group'], hint_level=q['hint_level']) for q in before['questions']])) evaluations = output / 'eval' if evaluations.exists() or evaluations.is_symlink(): if evaluations.resolve() != (folder / 'eval').resolve(): raise ValueError('Saved verification points to a different evaluation folder') else: evaluations.symlink_to(folder / 'eval', target_is_directory=True) write(output / 'result.json', paired(before, after)) value = read(output / 'result.json') low, high = value['uplift_ci95'] table.append(f'{case["model"].upper()}{case["task"]}' f'{value["sft_pass1"]:.1%}{value["rl_pass1"]:.1%}' f'[{100 * low:+.1f}, {100 * high:+.1f}] pp') return ('

Original held-out panels

Matched terminal SFT/RL evaluations; ' 'paired 95% gain intervals cluster source groups. The larger 32M verification is separate.

' '' + ''.join(table) + '
SizeTaskSFTRL95% gain interval
') if table else '' def update(root: Path, results: Path): plan = read(root / 'plan.json') baseline = read(ROOT / 'pilot-2m/plan.json') revised_gsm = plan['revision'] in ('gsm-visible-calculation-v2', 'gsm-final-step-reference-v3') records = ([] if revised_gsm else [summarize(case) for case in baseline['cases'] if case['model'] in ('8m', '32m')]) records += [summarize(case) for case in plan['cases']] figures = '' if revised_gsm: plot = runpy.run_path(str(Path(__file__).parent / '.plotting/visible.py'))['plot'] plot(root, results) figures = '

Fixed 32M diagnostic by missing length and final-calculation difficulty:

' if plan['revision'] == 'gsm-final-step-reference-v3': figures += saved_test_comparisons(root, plan) write(results / '.cache/comparison.json', records) fields = ['task', 'model', 'sft_tokens', 'recipe', 'sft_pass1', 'sft_hint1', 'rl_rollouts', 'rl_pass1', 'rl_hint1', 'complete', 'stopped', 'wandb'] rows = [] for record in records: sft, rl = record['sft'], record['rl'] rows.append(dict(task=record['task'], model=record['model'], sft_tokens=record['sft_tokens'], recipe=record['recipe'], sft_pass1=sft['metric']['pass1'] if sft else None, sft_hint1=sft['by_hint']['1']['pass1'] if sft else None, rl_rollouts=rl['counter'] if rl else 0, rl_pass1=rl['metric']['pass1'] if rl else None, rl_hint1=rl['by_hint']['1']['pass1'] if rl else None, complete=record['complete'], stopped=bool(record['stopped']), wandb=record['wandb'])) with (results / 'comparison.csv').open('w', newline='') as stream: writer = csv.DictWriter(stream, fieldnames=fields) writer.writeheader() writer.writerows(rows) table = [] for row in rows: cells = [] for key in fields[:-1]: value = row[key] if key in ('sft_pass1', 'sft_hint1', 'rl_pass1', 'rl_hint1'): value = '—' if value is None else f'{value:.1%}' cells.append('' + html.escape(str(value)) + '') table.append('' + ''.join(cells) + 'W&B') (results / 'index.html').write_text(''' Math optimizer controls

Math · SFT and RL controls

''' + html.escape(plan['controlled_comparison']) + '''

Scores use fixed development panels; each hint has ''' + str(plan.get('eval_queries', 64) // 16) + ''' questions. Stopped diagnostics are incomplete benchmarks.

CSV

''' + figures + '' + ''.join('' for key in fields) + '' + ''.join(table) + '
' + key + '
') @dataclass class ReportConfig: watch: bool = False interval: int = 60 publish: bool = False def refresh(config: ReportConfig): from rich import print while True: runpy.run_path(str(Path(__file__).parent / '.plotting/iterative.py'), run_name='__main__') if not config.watch: return validation = ROOT / 'verification/arithmetic-balanced-job.json' plan = ROOT / 'iterations/projection-only-v1/plan.json' ready = validation.exists() and plan.exists() if ready: ready = all((Path(case['output']) / 'result.json').exists() for case in read(validation)['cases']) ready = ready and all(any((Path(case['output']) / marker).exists() for marker in ['complete.json', 'stopped.json']) for case in read(plan)['cases']) auxiliary = ROOT / 'iterations/aux-smallmath-v1/plan.json' if auxiliary.exists(): ready = ready and all(any((Path(case['output']) / marker).exists() for marker in ['complete.json', 'stopped.json']) for case in read(auxiliary)['cases']) curriculum = ROOT / 'iterations/rl-smallcalc-v1/plan.json' if curriculum.exists(): ready = ready and all(any((Path(case['output']) / marker).exists() for marker in ['complete.json', 'stopped.json']) for case in read(curriculum)['cases']) for suite in ('gsm-visible-v2', 'gsm-final-step-v3'): experiment = ROOT / 'iterations' / suite / 'plan.json' if experiment.exists(): update(experiment.parent, Path(read(experiment)['results'])) ready = ready and all(any((Path(case['output']) / marker).exists() for marker in ['complete.json', 'stopped.json']) for case in read(experiment)['cases']) expanded = ROOT / 'verification/full-arithmetic-balanced-job.json' if expanded.exists(): ready = ready and all((Path(case['output']) / 'result.json').exists() for case in read(expanded)['cases']) print('[cyan]Math overview refreshed[/cyan]', 'complete' if ready else 'experiments running', flush=True) if ready: if config.publish: from .publish import Config as PublishConfig, run run(PublishConfig()) return time.sleep(min(60, max(1, config.interval))) if __name__ == '__main__': from omegaconf import OmegaConf refresh(OmegaConf.to_object(OmegaConf.merge(OmegaConf.structured(ReportConfig), OmegaConf.from_cli())))