klokedm/tabnetics-fsready — FS-ready episode shards (per-episode supervision)
klokedm/tabnetics-synt-ground-truth — world introspection (per-world generative-process metadata)

The data here supplies the clean population truth that the trainer uses as supervision for its population-reconstruction auxiliaries (population_reconstruction_targets, population_class_reconstruction_targets).

Repository layout

catalog_manifest.json                                Top-level v3 manifest
slices/
  expanded_126/
    population/train/w1024/                          (added 2026-05-24)
      stats/world=<world_id>/population_stats.npz
      class_moments/world=<world_id>/population_class_moments.npz
    population/train/w2048/                          (scheduled by bsc-run/w2048_backfill_guard.sh)
    worlds/train/lab{01,03}/synthetic/<archetype>/<world_id>/   (added 2026-05-24, all widths)
  gap/
    population/train/w1024/
      stats/world=<world_id>/population_stats.npz
      class_moments/world=<world_id>/population_class_moments.npz
    population/validation/w1024/
      stats/, class_moments/
  baseline/                                          migration_pending — via tabnetics-synt-full-v2
  unseen_validation/                                 planned

What each file contains

`population_stats.npz`

tabnetics_world_population_stats_v1 schema. Computed from clean row banks (no corruption). Used as target for the population_reconstruction_targets loss.

Array	Dtype / shape
`marginal_stats`	float32 [F, 5] (`mean`, `std`, `skew`, `kurt`, `median`)
`class_stats`	float32 [H, F, 5] (`fisher_log`, `mean_diff`, `std_diff`, `skew_diff`, `kurt_diff`)
`marginal_channels`, `class_conditional_channels`	str channel names
`world_id`, `archetype_id`, `width_bucket`, `row_bank_rows`, `feature_count`, `feature_ids`, `head_ids`, `head_class_counts`, `schema_version`	scalars / arrays

`population_class_moments.npz`

tabnetics_world_population_class_moments_v1 schema. Used as target for population_class_reconstruction_targets.

Array	Dtype / shape
`marginal_moments`	float32 [F, 5]
`marginal_quantiles`	float32 [F, 10]
`class_moments`	float32 [H, K, F, 5]
`class_quantiles`	float32 [H, K, F, 10]
`class_mean_rank`	float32 [H, K, F]
`class_n`, `class_present`	int64 [H, K] / bool [H, K]
`quantile_probabilities`	float32 [10]

`slices/expanded_126/worlds/train/lab{01,03}/synthetic/<archetype>/<world_id>/`

The clean row-bank world bundles (metadata.json, parquet row banks, head definitions). 1920 files / 13.9 GB across both labs and all widths.

Coverage matrix (w1024 closed 2026-05-24; w2048 scheduled)

Slice	worlds	files	bytes	availability
`expanded_126:population_stats:train:w1024`	48	97	~35 MB	available (added 2026-05-24)
`expanded_126:population_class_moments:train:w1024`	48	97	~2.5 GB	available (added 2026-05-24)
`expanded_126:worlds:train:w1024-w8192`	192	1920	13.9 GB	available (added 2026-05-24)
`expanded_126:population_stats:train:w2048`	(192 expected)	TBD	TBD	scheduled by `bsc-run/w2048_backfill_guard.sh`
`expanded_126:population_class_moments:train:w2048`	(192 expected)	TBD	TBD	scheduled
`expanded_126:worlds:validation:w1024-w8192`	—	—	—	planned (no expanded_126 validation episodes yet)
`gap:population_stats:train:w1024`	30	63	~12.5 MB	available (consolidated 2026-05-10)
`gap:population_class_moments:train:w1024`	30	62	~0.7 GB	available
`gap:population_stats:validation:w1024`	30	63	~12.5 MB	available
`gap:population_class_moments:validation:w1024`	30	62	~0.7 GB	available
`baseline:fsready:train:w1024:quota`, `baseline:episodes:train:w1024`, `baseline:worlds:all:w1024-w8192`	—	—	—	migration_pending (use `tabnetics-synt-full-v2`)
`baseline_v1:worlds:all:w1024-w8192`	—	—	—	source_pending

License

Apache-2.0.

Downloads last month: 3

Total file size:

34.4 GB