README
VCDesign-CED processed data
This repository stores the frozen processed inputs used by the VCDesign-CED paper, plus the selected base-model checkpoint and compact evaluation records. The source single-cell H5AD files are obtained from the original public releases and are not duplicated here.
Contents
| Directory | Contents |
|---|---|
inputs/ | The 19 frozen inputs referenced by the paper's training configuration, plus source-goal metadata |
inputs/four_context/packs/ | K562, RPE1, HepG2 and Jurkat processed context packs |
inputs/external_baselines/scores/ | Frozen external-comparator score matrices for RPE1, HepG2 and Jurkat |
checkpoints/ | Selected epoch-8 VCDesign-CED base-model checkpoint |
records/ | Training, selection, evaluation, context-build and comparator verification records |
code/ | Versioned, anonymous code-package ZIP for local download and review |
DATA_MANIFEST.json | File sizes, release SHA-256 values and original SHA-256 values where a record was anonymized |
Code package
Download the latest verified code package directly:
VCDesign_ICLR2027_code_v0.2.0.zip
SHA-256: 78c96552727b378ecde418bac9e4f9015739c5e49e4aeac7b57a345c220c40ab
Version 0.2.0 places gene_open_inverse/ directly at the package root and adds
an anonymous, visual README with a complete local copy of Figure 1. The ZIP also
contains the frozen configuration, exact CPU verification dependencies, a
synthetic demo, 162 protocol tests, and its own per-file
CODE_PACKAGE_MANIFEST.sha256. It does not contain large processed inputs,
checkpoints, Git history, or author-account links.
Previous release: v0.1.0
DATA_MANIFEST.json identifies the exact bytes uploaded. JSON records that
contained compute-host paths have only those path strings replaced with release
relative paths or UNBUNDLED/<filename> labels. The original file hash is
retained beside each transformed record's release hash. Numeric arrays and model
weights are copied byte-for-byte and checked against the frozen run manifest.
The processed data correspond to the code package's
configs/paper_run_v1.json path layout. To place only the data files in
a checkout of that package without replacing its README:
hf download Boom5426/VCDesign --repo-type dataset \
--include 'inputs/**' 'checkpoints/**' 'records/**' 'DATA_MANIFEST.json' \
--local-dir /path/to/code-package
python3 tools/verify_processed_data.py --root /path/to/code-package
The exact repository revision should be pinned for a reproduction run. The
selected checkpoint was epoch 8, with frozen SHA-256
58da13d252faf9bcaa74c53345c0afc380882bdd5f410a77cfa5044f9a6e9484.
Original public sources
- Replogle et al. K562 genome-scale and RPE1 Perturb-seq
- GEO GSE264667 HepG2 and Jurkat screens
- STRING v12.0 physical interaction data
- MAP and MAP-KG
- ESM-2
The derived files are tied to those source versions; the K562 genome-scale screen is distinct from the K562 essential screen. Cite the original data and knowledge resources when using this collection.
Similar repos
README
VCDesign-CED processed data
This repository stores the frozen processed inputs used by the VCDesign-CED paper, plus the selected base-model checkpoint and compact evaluation records. The source single-cell H5AD files are obtained from the original public releases and are not duplicated here.
Contents
| Directory | Contents |
|---|---|
inputs/ | The 19 frozen inputs referenced by the paper's training configuration, plus source-goal metadata |
inputs/four_context/packs/ | K562, RPE1, HepG2 and Jurkat processed context packs |
inputs/external_baselines/scores/ | Frozen external-comparator score matrices for RPE1, HepG2 and Jurkat |
checkpoints/ | Selected epoch-8 VCDesign-CED base-model checkpoint |
records/ | Training, selection, evaluation, context-build and comparator verification records |
code/ | Versioned, anonymous code-package ZIP for local download and review |
DATA_MANIFEST.json | File sizes, release SHA-256 values and original SHA-256 values where a record was anonymized |
Code package
Download the latest verified code package directly:
VCDesign_ICLR2027_code_v0.2.0.zip
SHA-256: 78c96552727b378ecde418bac9e4f9015739c5e49e4aeac7b57a345c220c40ab
Version 0.2.0 places gene_open_inverse/ directly at the package root and adds
an anonymous, visual README with a complete local copy of Figure 1. The ZIP also
contains the frozen configuration, exact CPU verification dependencies, a
synthetic demo, 162 protocol tests, and its own per-file
CODE_PACKAGE_MANIFEST.sha256. It does not contain large processed inputs,
checkpoints, Git history, or author-account links.
Previous release: v0.1.0
DATA_MANIFEST.json identifies the exact bytes uploaded. JSON records that
contained compute-host paths have only those path strings replaced with release
relative paths or UNBUNDLED/<filename> labels. The original file hash is
retained beside each transformed record's release hash. Numeric arrays and model
weights are copied byte-for-byte and checked against the frozen run manifest.
The processed data correspond to the code package's
configs/paper_run_v1.json path layout. To place only the data files in
a checkout of that package without replacing its README:
hf download Boom5426/VCDesign --repo-type dataset \
--include 'inputs/**' 'checkpoints/**' 'records/**' 'DATA_MANIFEST.json' \
--local-dir /path/to/code-package
python3 tools/verify_processed_data.py --root /path/to/code-package
The exact repository revision should be pinned for a reproduction run. The
selected checkpoint was epoch 8, with frozen SHA-256
58da13d252faf9bcaa74c53345c0afc380882bdd5f410a77cfa5044f9a6e9484.
Original public sources
- Replogle et al. K562 genome-scale and RPE1 Perturb-seq
- GEO GSE264667 HepG2 and Jurkat screens
- STRING v12.0 physical interaction data
- MAP and MAP-KG
- ESM-2
The derived files are tied to those source versions; the K562 genome-scale screen is distinct from the K562 essential screen. Cite the original data and knowledge resources when using this collection.