cached_stage_1_prior -- portable TabICLv2 stage-1 data bundle

Pre-generated stage-1 prior batches and a held-out evaluation set, packaged for copying to other Alliance clusters. Self-contained: no symlinks, no dependency on anything outside this directory.

Contents

path	contents
stage1_prior_v2/stream_0..7/	200 shards x 250 = 50,000 training batches, seeds 42-49
stage1_eval_v2/stream_0/	1 shard = 200 held-out batches, seed 999

Total: 643 GB apparent, ~285 GB on disk (the shards are sparse -- see below).

Each batch is 64 datasets: 1,024 rows each, 1-100 features, up to 10 classes, 30-90% train split, graph_scm prior, batch_size_per_gp=4.

state_*.pkl are RNG snapshots used to resume generation; they are not needed for training and can be dropped if space is tight (they occupy ~0 GB).

Provenance

Generated by pretraining/pregen_stage1.py with the prior configuration of scripts/train_v2_clf_stage1.sh (soda-inria/tabicl @ 46b9196) -- in particular min_features 1, max_features 100, batch_size_per_gp 4. Equality with the official prior was verified by constructing both PriorDataset objects and diffing every attribute plus all 27 PriorConfig fields: no differences.

Streams 0-7 are the first 8 of a 32-stream, 200,000-batch cache. The 50K-step results in pretraining/RESULTS.md were trained on exactly these 8 streams. The eval set is seed 999, outside the training seed range.

Copying to another cluster

Preserve sparseness or the bundle doubles in size on arrival:

rsync -aS --info=progress2 cached_stage_1_prior/ <dest>/cached_stage_1_prior/

Globus does not follow symlinks, which is why this directory holds real files rather than links back into stage1_prior_v2/.

Verify after transfer -- shard count and apparent size must match:

find <dest>/cached_stage_1_prior -name 'shard_*.pt' | wc -l     # expect 201
du -sh --apparent-size <dest>/cached_stage_1_prior              # expect ~647 GB
du -sh <dest>/cached_stage_1_prior                              # ~294 GB if sparse survived

Integrity

MANIFEST.sha256 holds a SHA-256 for all 402 files. It is the authoritative check and needs no reference copy, so it works on every destination cluster:

cd <dest>/cached_stage_1_prior && sha256sum -c MANIFEST.sha256

Expect OK on every line. Anything else means a corrupt or truncated file -- re-run the rsync -aS above, which re-transfers only the files that differ.

Reference hashes

Anchor -- verifies the manifest itself has not been altered:

a7591215e4d12370ee11712649eba49456dd1d2901d1e4bd7e98f5dfbfe196b4  MANIFEST.sha256

Per-stream rollups. Each is the SHA-256 of that stream's per-file hashes, concatenated in sorted-path order -- so one value covers the whole stream:

stream	files	rollup SHA-256
stream_0	50	db11725965bed60f52f9aedd8d73bfaf46204381777ee9a28ca1f18481107b85
stream_1	50	369fa625b51f52b616a2ebd073d8ef31e80fe17c70627b569d509fd0e4cf94b3
stream_2	50	3a19ba73fccfd155f471951b840d9404cfbfc849d130af21bc87d71b3a9b94c5
stream_3	50	df1991a717f0b0c29f999ab714b175df8ad5edb703437f407af34cede0e3acd4
stream_4	50	99fc74f569fa467cba933a46a65e3f0b0ec0b1097956cc4231cf924f1286473d
stream_5	50	4cbd3528c3110844ce9b2069952b5c95798873946afa71ff79825000acfdc335
stream_6	50	6c36d36581e366a728ada8c03590ba830769696d01f4b9cc74539bdf1887dd44
stream_7	50	6bccc7334eed832651f1b44d2f6f1a8e70c923de8490484d7336922a4c9e3782
stage1_eval_v2	2	8695e1e8d9a5c652a76a9c43ea09a44c79f9b9526f4c3047bdcb14f1c4a55fe5

402 files total (200 shards + 200 state_*.pkl + 1 eval shard + 1 eval state). Reproduce any row with:

grep " stage1_prior_v2/stream_N/" MANIFEST.sha256 | sort -k2 | awk '{print $1}' | sha256sum

This is a per-stream spot check, useful when a transfer of one stream is suspect. sha256sum -c MANIFEST.sha256 remains the complete check.

For a faster smoke test on arrival (checks truncation, not bit-rot):

find . -name 'shard_*.pt' | wc -l          # 201
du -sh --apparent-size .                   # ~647 GB

Note that rsync writes to a temporary name and renames only on success, so an interrupted transfer leaves a hidden .shard_*.pt.XXXXXX orphan rather than a truncated shard. If a transfer is killed, delete any such orphans and re-run:

find . -name '.*.pt.*' -delete

Using it

python pretraining/stage1_train.py --data_dir <path>/cached_stage_1_prior/stage1_prior_v2 ...

and point EVAL_DIR in pretraining/eval_arms.py at <path>/cached_stage_1_prior/stage1_eval_v2/stream_0.

Shards are read in sorted glob order and concatenated, so the 8 streams behave as one 50,000-batch stream.

This bundle supports at most 50,000 training steps. DiskPrefetcher walks its shard list once and then signals exhaustion -- it does not wrap. A longer run needs either more streams or a cycling loop added to disk_loader.py; repeating data would in any case be a deviation from the reference, which draws 200,000 distinct batches over a 200K-step run.