VIEWER_MIGRATION.md

Dataset Viewer Migration

Goal

Expose canonical and pre-standard recordings as separate Hub configurations so users do not mistake metadata-incomplete historical material for canonical standard data.

Why this is not applied in v0.2.0

The current repository places canonical MP4 files directly in Standard_Time_Lapses/ and supporting provenance holdings beneath Standard_Time_Lapses/Provenance/. A broad data_dir configuration could therefore ingest noncanonical provenance videos.

Hugging Face supports multiple configurations and folder-local video metadata, but the exact direct-file glob and metadata discovery behavior should be tested through a Hub pull request before changing the public default configuration.

Proposed pull-request test

1. Add a folder-local canonical metadata file whose file_name values are relative to Standard_Time_Lapses/.
2. Add a folder-local pre-standard metadata file with explicit collection_tier = pre_standard and is_canonical = false.
3. Add two README configurations:
   • canonical — default, direct MP4 files only
   • pre_standard — separate historical subset
4. Use a direct-file glob for canonical MP4 files so the Provenance/ subtree is excluded.
5. Confirm the PR preview reports the exact active-index counts, currently 338 canonical and 129 pre-standard rows, with zero provenance videos and zero unindexed direct files.
6. Confirm folder-local metadata is attached to every expected row.
7. Merge only after Viewer and Parquet generation succeed.

Candidate YAML for PR testing

This block is a test proposal, not the current card configuration:

configs:
- config_name: canonical
  default: true
  data_dir: Standard_Time_Lapses
  data_files:
  - split: archive
    path: "*.mp4"
  drop_labels: true
- config_name: pre_standard
  data_dir: Pre_Standard_Time_Lapses
  data_files:
  - split: archive
    path: "*.mp4"
  drop_labels: true

If folder-local metadata is not discovered with the direct-file glob, do not broaden the glob. Instead, create explicit Parquet manifests for the two configurations so row membership is unambiguous.

Split terminology

Use archive rather than train if supported by the Viewer build. Neither configuration should imply an official machine-learning evaluation split.