EnriFermi/weightclip-resnet18slim-operator-bank
WeightCLIP ResNet18Slim operator bank This private dataset repository stores the exact preprocessed operator dataset used by the current WeightCLIP production training run. It contains the active weight-tile bank, the matching activation-context bank, and the metadata needed to reproduce production sampling. The active payload is 41,724,270,793 bytes (38.859 GiB) across 3,622 source files: weight_tile_bank-9020c363ee2005e8: 135,100 weight records, 793 files, 11,181,671,316… See the full description on the dataset page: https://huggingface.co/datasets/EnriFermi/weightclip-resnet18slim-operator-bank.
WeightCLIP ResNet18Slim operator bank
This private dataset repository stores the exact preprocessed operator dataset used by the current WeightCLIP production training run. It contains the active weight-tile bank, the matching activation-context bank, and the metadata needed to reproduce production sampling.
The active payload is 41,724,270,793 bytes (38.859 GiB) across 3,622 source files:
weight_tile_bank-9020c363ee2005e8: 135,100 weight records, 793 files, 11,181,671,316 bytes;context_bank-234e31b2ec93af4f: 90,300 activation-context records, 2,825 files, 30,517,934,053 bytes;operator_dataset-c478efe716506765.json, coverage metadata, and both permutation-view files.
This snapshot deliberately excludes older bank generations, raw source model checkpoints, and cached experiment outputs. The authoritative machine-readable contract is dataset_contract.json in the Hub repository and conf/weightclip_benchmark/operator_dataset_remote_hf.json in the Git source.
Download on another machine
The repository is private. Authenticate with a Hugging Face token that has read access, then run from the Git checkout:
export HF_TOKEN=...
python projects/weight-vae/workspace/scripts/download_weightclip_operator_dataset_from_hf.py \
--output-root /data/weightclip_resnet18slim_operator_datasetThe downloader pins the revision recorded in the Git contract, validates the small immutable manifests and expected file inventory, and writes a resolved pair manifest whose runtime paths point into the downloaded snapshot. It prints the resolved manifest path and its SHA-256; use those values for operator_bank.pair_manifest and operator_bank.pair_manifest_sha256 in a local production config.
The original pair manifest is retained byte-for-byte. It contains absolute paths from the machine that built the dataset. Only its top-level runtime paths need relocation; absolute paths nested under contract are historical provenance and are not required to read the preprocessed banks.
Source contract
- Pair manifest SHA-256:
f3a26fc9772d43feabe138ef9fa5be5baf2f76323fb4f3079c12c5f65275f347 - Dataset contract SHA-256:
c478efe716506765af35236942e128d416472690d6dfc2cb5045bdfc359b562e - Weight-bank content SHA-256:
9020c363ee2005e88c5a9bc746e97c8620cf62c6fac94f4ed865b08bb7d9a19a - Context-bank content SHA-256:
234e31b2ec93af4f3bce294a5c09c202ae5c560d0faf2f5515c00ddfc3104bdd
The bank content hashes above were sealed when the banks were built. The download helper uses the small manifest hashes plus file inventory and does not rehash all 38.9 GiB. The current production OperatorDatasetBank loader is stricter: when it opens the materialized banks, it SHA-256-validates every payload file. Plan for one full sequential read during first dataset open on a new machine.
