klokedm/tabnetics-synt-full-v2

Tabnetics Synthetic Worlds Full V2

This README describes the tabnetics-synt-full-v2 synthetic-world catalog as an all-in trainer source-of-truth repo (klokedm/tabnetics-synt-full-v2) and, where mirrored, as a versioned slice in klokedm/tabnetics-synthetic-worlds.

Catalog Identity

• Catalog key: full-v2
• Catalog name: tabnetics-synt-full-v2
• Canonical slice id: baseline:worlds:all:w1024-w8192
• Consolidated storage path: slices/baseline_v2/worlds/all/w1024-w8192
• All-in trainer repo: klokedm/tabnetics-synt-full-v2
• Training contract artifact: catalog/training_contracts/synthetic_v2_training_contract.json
• Availability: migration_pending
• Role: Current pre-#126 baseline synthetic-world catalog used by existing HDLSS episode and FS-ready bundles.

Shape

• Worlds: 120
• Archetypes: 10
• Width buckets: 1024, 2048, 4096, 8192
• Seeds per archetype/width: 0, 1, 2
• Task heads per world: 20
• Total task heads: 2400

Consolidated Layout

The catalog is published under the lifecycle-stage repository tabnetics-synthetic-worlds, not as a standalone dated repo. Each world remains a structured directory under the slice path with clean row banks, metadata, generator state, labels, feature stats, and ground truth.

Expected root-level files for the staged slice:

• README.md
• manifest.json
• catalog_version.json
• ground_truth_index.json
• holdout_gates.json
• training_contract.json in packaged consolidated repos
• selectors/{all,train,validation}/all.json
• teacher_meta_features.parquet and teacher_meta_features_summary.json
• sidecars/population_stats/... and sidecars/population_class_moments/...
• synthetic/<archetype>/<world_id>/...

Training Row Roles

Same-world train-time validation rows are drawn from each training world, disjoint from sample and classifier-OOS rows; the validation package may also include existing extra held-out validation worlds. Every FS-ready shard derived from this catalog must carry support/sample rows, 200 classifier out-of-sample rows from the same world, and train-time validation rows from the same world. These row groups must be mutually disjoint within each episode.

Reproducibility

• Manifest schema version: 3
• Migration/publish command template: python -m scripts.training.publish_hf_catalog_slices --slice-id baseline:worlds:all:w1024-w8192 --download-legacy ...
• Source refs: [{"legacy_repo_id": "klokedm/tabnetics-synt", "legacy_role": "full-v2 worlds"}]

Additive V3 Sidecars

The world rows, feature IDs, task heads, labels, and generator metadata are preserved exactly from the source catalog. The v3 consolidation only adds derived sidecars for training-time statistics:

• population_stats.npz from scripts/training/precompute_world_population_stats.py.
• population_class_moments.npz from scripts/training/precompute_world_population_class_moments.py.
• FS-ready sample/observed sidecars are generated in the tabnetics-fsready lifecycle repo with scripts/training/precompute_fsready_sidecars_v2.py.
• FS-ready target-v2 addenda are generated with scripts/training/precompute_fsready_targets_v2.py.
• Promotion requires scripts/training/verify_v3_sidecar_stats.py to confirm the new population and sample statistics are present.
• Promotion also requires scripts/training/verify_synthetic_world_repo_contract.py to confirm ground truth, labels, row banks, selectors, holdout gates, teacher meta-features, and train/validation split metadata are present.

Status

Migration-pending consolidated slice sourced from the legacy full-v2 HF repo.

Promotion to available requires content-bound slice signatures, file counts, byte totals, and the catalog manifest update to be verified by the #143 integrity tooling.