muhammadhasyim/tps-cofolding-paper-data

TPS cofolding research data

Scientific outputs and frozen analysis inputs for TPS cofolding, including protein and ligand sampling campaigns, checkpoint histories, structural coordinates, collective-variable histories, and paper figure inputs.

Find the data

The existing paths are preserved so that manuscript and analysis links keep working. Data have not been moved or renamed by this catalog update.

Directory	Contents
outputs/ablation_5x11_af3/	Main sampling campaigns, saved ensembles, and checkpoints
outputs/ablation_5x11_af3_nofreeze/	No-freeze comparison outputs
outputs/ablation_5x11_af3_barrier250/	Barrier-setting comparison outputs
outputs/eftu_switch_af3/	Frozen EF-Tu context outputs used in analysis
outputs/rnp_duality_af3_exact/	Runs N' Poses protein–ligand benchmark outputs
outputs/analysis/	Frozen numerical analyses and figure ingredients
data/	Campaign definitions and external reference inputs
cache/rcsb/	Reference structure files
catalog/	Versioned inventories and storage/restore documentation
tools/	Download verification and restore helper

Inventory and versions

The inventory captured on 26 October 2026 records 6,590 files and 2.294 TB of logical file content at commit 5a9423b68c3c. Logical size is not a measure of deduplicated backend storage.

• Full file inventory: paths, byte counts, Git blob IDs, and LFS SHA-256 digests where available.
• Inventory summary: source commit and size breakdown.
• manifest.json is the original curated subset manifest, not a full inventory of all later uploads.

The repository includes multiple frozen campaign checkpoints and analysis dates. A directory name alone does not establish one common scientific snapshot. Select an exact checkpoint and pin the dataset commit when reproducing an analysis.

Download selectively

Download individual files rather than the entire multi-terabyte repository:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="muhammadhasyim/tps-cofolding-paper-data",
    repo_type="dataset",
    revision="5a9423b68c3cc9b3eadec7672a7bc245e714b6be",
    filename="manifest.json",  # replace with a path from the full inventory
    local_dir="./paper-data",
)

For large files, verify the downloaded byte count and SHA-256 against the inventory. Small ordinary Git files use Git blob IDs in the inventory; these include the Git blob header and are not ordinary file SHA-1 digests.

Local archive and restore receipts

tools/hf_archive.py verifies a local file against an immutable Hub commit and can perform a complete independent download and checksum comparison. It retains the local source during verification. A local .hf-archive.json receipt records an evicted file's exact repository, commit, path, size, and SHA-256. Restore one with an explicit destination:

python tools/hf_archive.py restore /path/to/file.npz.hf-archive.json /path/to/file.npz

The helper requires Python 3.10+ and huggingface_hub. Its eviction command is restricted to the original project's frozen non-CloudRift ablation exports; it does not manage live jobs or CloudRift's operational backup repository.

Operational boundary

This is the public scientific-data repository. Live CloudRift backup ownership, retention, and runtime health are managed separately. A file's presence here is not permission to remove a live checkpoint, a required recovery generation, or a reconstruction dependency.

Verified cold research outputs (2026-10-26)

Immutable inactive scientific payloads are organized under archives/2026-10-26/repo-cold/objects/ (large individual files) and archives/2026-10-26/repo-cold/packs/ (small-file tar shards). Recovery maps under catalog/cold-storage/2026-10-26/ record pinned data revisions, payload lengths, SHA256 digests and original relative-path aliases. The final index.json is published when the transfer workers finish.

Local copies are removed only after a complete independent remote download and SHA256 verification. Tar shards additionally verify every member. Active production cells and CloudRift recovery data are excluded from this cold tier. Archived checkpoints retain their exact bytes and completed rounds and must be restored before resuming their original runs.

Download tools/restore_cold.py and a recovery map (at its pinned map revision), install huggingface_hub, then run:

python restore_cold.py recovery-map.json --project-root /path/to/tps-cofolding

Add --staging-root /path/to/for_paper to restore upload-staging hard-link aliases; use --member ID for a single member of a tar shard. Existing different files are never overwritten. Recovery maps and archived data remain separate from the legacy dataset paths used by existing paper scripts.

Closed CloudRift checkpoint backups (2026-10-28)

Five closed checkpoint stores (workers 01, 02, 03, 04 and 07) preserve all 99 older states in 137 immutable tar shards. The archive includes 288,658,631,983 original file bytes, all state indices, frozen restore tools, the source catalog and restoration references. After verification, 286,810,295,604 blob payload bytes were removed from Torch; indices, tools, original sources, newest two verified generations per cell, structures, journals and final readers remain.

Every shard/member passed a fresh-download SHA256 check, each compressed blob was decoded and matched to its content digest, and one full checkpoint from each store was reconstructed from remote-only inputs against its original length and SHA256. The published independent restore helper also passed a real restore into an empty temporary project.

Checkpoint backup and recovery index. The index pins each worker recovery map; worker maps contain all shard maps and the exact helper revision/checksum in their source metadata. Run the frozen restore_cold.py helper with a shard map and --project-root to restore its files, or add --member from an adjacent local receipt to restore one payload. Restore archived blobs first, then use the unchanged frozen state restore tool and state index to reconstruct the original state.pt. Existing unequal files are never overwritten.