Workflow_Subsets/README.md

Standard workflow selections

The two CSVs in this folder are generated access views of the authoritative root Standard index. Membership uses only the existing acquisition_protocol_version values:

• standard_earlier: standard-2025-07-13.
• standard_individual_recordings: standard-2026-08-27.

Each CSV preserves the source header, row order, and every field value. Its file_name paths are relative to the repository root, not this folder. The native configurations use the root index, an explicit protocol filter, and only the selected media paths. They retain the existing typed video feature and archive split. Unknown starts and unresolved work IDs remain included. The two views partition canonical; do not add their counts to it.

Maintain the generated views

Regenerate after the normal identity/timing projections and canonical count update, following the existing ingest procedure:

python tools/build_workflow_subsets.py --root . --write
python tools/build_workflow_subsets.py --root . --check-loading
python -m unittest discover -s tools -p 'test_*.py'

The repository pins datasets==4.8.5 for loading; generation also needs PyYAML. Counts, CSV views, metadata filters and explicit media-path lists are generated together; do not edit generated outputs manually. The loading check requires the complete repository paths (Git LFS pointers suffice), refuses video payload reads, checks exact selected input paths, and compares every loaded field to the source. This does not establish hosted Viewer cache readiness. Regeneration is part of the manual maintainer procedure; it is not a scheduled or automated ingestion service.

Retain a selection before downloading videos

Resolve the desired revision once and save its full commit and chosen IDs. This example downloads only the generated CSV; media downloads remain separate. Use a revision containing the configurations, since earlier immutable releases retain their original layout.

import csv
import hashlib
import json
from pathlib import Path
from huggingface_hub import HfApi, hf_hub_download

repo = 'maxwellinked/time-lapse-artifacts'
config = 'standard_earlier'  # or 'standard_individual_recordings'
revision = HfApi().dataset_info(repo, revision='main').sha
index_name = f'Workflow_Subsets/{config}.csv'
index = Path(hf_hub_download(repo, index_name, repo_type='dataset', revision=revision))
with index.open(encoding='utf-8', newline='') as handle:
    rows = list(csv.DictReader(handle))

# Optionally filter timing evidence, dates, materials, or confirmed work IDs
# here. Retain the filter description and the exact resulting IDs.
selection = {
    'repo_id': repo,
    'revision': revision,
    'config': config,
    'split': 'archive',
    'index_path': index_name,
    'index_sha256': hashlib.sha256(index.read_bytes()).hexdigest(),
    'filters': [],
    'record_ids': [row['record_id'] for row in rows],
}
Path('selection.json').write_text(json.dumps(selection, indent=2) + '\n', encoding='utf-8')

Load the same unfiltered workflow through the native library at that revision:

from datasets import Video, load_dataset

records = load_dataset(repo, config, split='archive', revision=revision, streaming=True)
records = records.cast_column('video', Video(decode=False))
assert {row['record_id'] for row in records} == {row['record_id'] for row in rows}

If optional filters were applied to rows, apply the same filters to the loaded records before this comparison. For an intentional media download, use a retained record ID and commit with the existing verified downloader:

python tools/download_verified.py RECORD_ID --revision FULL_COMMIT_SHA --check-only
# Replace --check-only with --output-dir DESTINATION for the selected download.

The saved selection records a research population. It does not create new archive classifications, work identities, or timing observations.