| # Dataset Viewer Migration |
|
|
| ## Goal |
|
|
| Expose canonical and Pre-Standard recordings as separate Hub configurations so |
| users do not mistake different acquisition/documentation eras for an official |
| machine-learning split. |
|
|
| ## Current configuration |
|
|
| The release candidate defines: |
|
|
| - `canonical/archive` — default; 356 direct Standard MP4 files |
| - `pre_standard/archive` — 129 direct Pre-Standard MP4 files |
|
|
| Both configurations use explicit direct-file globs. The canonical glob does |
| not recurse into `Standard_Time_Lapses/Provenance/`, so supporting historical |
| videos cannot enter the canonical Viewer accidentally. |
|
|
| ```yaml |
| configs: |
| - config_name: canonical |
| default: true |
| data_files: |
| - split: archive |
| path: |
| - metadata.csv |
| - Standard_Time_Lapses/*.mp4 |
| - config_name: pre_standard |
| data_files: |
| - split: archive |
| path: |
| - Pre_Standard_Time_Lapses/metadata.csv |
| - Pre_Standard_Time_Lapses/*.mp4 |
| ``` |
|
|
| The two configurations share the same declared 29-field schema. Both tiers now |
| have complete core encoded-video metadata, stable IDs, content hashes, and byte |
| sizes. Historical descriptive and provenance uncertainty remains explicit in |
| the Pre-Standard tier. |
|
|
| ## Merge criteria |
|
|
| Merge only after the pull-request preview confirms: |
|
|
| 1. `canonical/archive` contains exactly 356 rows. |
| 2. `pre_standard/archive` contains exactly 129 rows. |
| 3. No provenance video is present in either configuration. |
| 4. Every expected row contains its metadata. |
| 5. Stored dimensions and byte sizes load as integers. |
| 6. `has_audio` loads as a boolean. |
| 7. Generated Parquet artifacts succeed for both configurations. |
|
|
| ## Split terminology |
|
|
| `archive` describes repository membership. It is not a training or evaluation |
| recommendation. No official train/validation/test partition exists. |
|
|
