| """Rebuild Endless v3 tasks.parquet and archive.parquet from pinned v2 Parquets and metadata/selection.json. |
| |
| pip install pyarrow huggingface_hub |
| python patch.py --out rebuilt |
| Optional --source-dir holds v2 tasks.parquet and archive.parquet for offline reproduction. |
| """ |
| import argparse, json |
| from pathlib import Path |
| import pyarrow as pa |
| import pyarrow.parquet as pq |
|
|
|
|
| def build(out, source_dir=None): |
| m = json.loads((Path(__file__).parent / 'metadata/selection.json').read_text()) |
|
|
| def source(name): |
| if source_dir: |
| return source_dir / name |
| from huggingface_hub import hf_hub_download |
| return Path(hf_hub_download(m['repo_id'], f'endless_v2/{name}', repo_type='dataset', |
| revision=m['v2_revision'], cache_dir=out / '.source-cache')) |
| out.mkdir(parents=True, exist_ok=True) |
| assert not any((out / n).exists() for n in ['tasks.parquet', 'archive.parquet']) |
| old, arch = pq.read_table(source('tasks.parquet')), pq.read_table(source('archive.parquet')) |
| extra = [pa.field('qwen38_attempts', pa.int64()), pa.field('qwen38_successes', pa.int64()), |
| pa.field('qwen38_target_attempts', pa.int64()), pa.field('kept_from_v2', pa.bool_())] |
| kept, excluded = [], [] |
| for r in old.to_pylist(): |
| assert m['tasks'][r['path']]['kept_from_v2'] |
| kept.append({**r, 'qwen38_attempts': 0, 'qwen38_successes': None, 'qwen38_target_attempts': None, 'kept_from_v2': True}) |
| for r in arch.to_pylist(): |
| e = m['tasks'][r['path']] |
| assert r['content_sha256'] == e['qwen38_evaluated_content_sha256'] |
| r.update(qwen38_attempts=e['qwen38_attempts'], qwen38_successes=e['qwen38_successes'], |
| qwen38_target_attempts=m['attempts'], kept_from_v2=False) |
| if e['kept']: |
| r.update(archive_stage=None, archive_reason=None, archive_labels=[]) |
| kept.append(r) |
| else: |
| r.update(archive_reason='zero_success_three_models', archive_labels=['zero_success_three_models']) |
| excluded.append(r) |
|
|
| def images(rows): |
| merged = json.loads(old.schema.metadata[b'ot.images.v1']) |
| other = json.loads(arch.schema.metadata[b'ot.images.v1']) |
| merged['tasks'].update(other['tasks']) |
| merged['bundles'].update(other['bundles']) |
| names = {r['path'] for r in rows} |
| merged['tasks'] = {t: e for t, e in merged['tasks'].items() if t in names} |
| used = {e['bundle'] for t in merged['tasks'].values() for e in t['environments']} |
| merged['bundles'] = {k: v for k, v in merged['bundles'].items() if k in used} |
| if 'uncached_task_ids' in merged: |
| merged['uncached_task_ids'] = sorted(names - set(merged['tasks'])) |
| merged['coverage'] = {'cached_tasks': len(merged['tasks']), 'total_tasks': len(names)} |
| return json.dumps(merged).encode() |
| s = m['summary'] |
| release = json.dumps({k: s[k] for k in ['previous_revision', 'source_folder', 'source_revision', 'new_model', 'attempts', |
| 'evaluated', 'v2_kept', 'newly_solved', 'kept', 'reward_policy']}).encode() |
| for name, table, rows in [('tasks.parquet', old, kept), ('archive.parquet', arch, excluded)]: |
| schema = pa.schema([*table.schema, *extra]) |
| pq.write_table(pa.Table.from_pylist(sorted(rows, key=lambda r: r['path']), schema=schema) |
| .replace_schema_metadata({b'ot.images.v1': images(rows), b'ot.release.v3': release}), |
| out / name, compression='zstd') |
| assert len(kept) == s['kept'] and len(excluded) == s['excluded'] |
| print(json.dumps({'kept': len(kept), 'excluded': len(excluded)})) |
|
|
|
|
| if __name__ == '__main__': |
| p = argparse.ArgumentParser(description=__doc__) |
| p.add_argument('--out', type=Path, required=True) |
| p.add_argument('--source-dir', type=Path) |
| a = p.parse_args() |
| build(a.out, a.source_dir) |
|
|
