CoolFace
Datasetpublic

ekunish/parc2026-track2-network-archive-20260912

Track 2 experiment recovery archive This repository preserves deduplicated data and numerical evidence from the owner's completed PARC 2026 Track 2 experiments. It is not a new training dataset release or evidence of model improvement. Original source licenses and attributions remain applicable to bundled code/assets. The migration may still be in progress. A volume's RECOVERY_CATALOG.json is published only after all its archives, original-path mappings, quarantine contents and… See the full description on the dataset page: https://huggingface.co/datasets/ekunish/parc2026-track2-network-archive-20260912.

sourceHugging Faceupdated 11d agoView on Hugging Face
0likes525downloads
Dataset Card

Track 2 experiment recovery archive

This repository preserves deduplicated data and numerical evidence from the owner's completed PARC 2026 Track 2 experiments. It is not a new training dataset release or evidence of model improvement. Original source licenses and attributions remain applicable to bundled code/assets.

The migration may still be in progress. A volume's RECOVERY_CATALOG.json is published only after all its archives, original-path mappings, quarantine contents and associated checkpoint receipts have been verified. A catalog does not, by itself, certify that the original RunPod reservation was deleted.

Contents

  • —nqupo8py43/: historical SmolVLA/Kairos teacher and evaluation artifacts.
  • —1lt1si5pwn/: capacity-experiment evaluation results and runtime metadata.
  • —Each volume has archives/, indices/ and a NONREGULAR.json manifest.
  • —Archives store each unique body once as objects/<SHA256>. Index entries map it back to every original path, including duplicate paths.
  • —Early full Kairos checkpoints are preserved separately at the capacity model repository.

Restore a batch

Install huggingface_hub, then obtain restore_backup-v2.py and an index at the immutable commit recorded in the volume's recovery catalog. The helper verifies the archive hash, every member hash, every restored file hash, file sizes and path safety. It never executes bundled code or loads pickle/model contents.

sh
python restore_backup-v2.py archive batch-0000-INDEX.json new-output-directory

The output directory must not already exist. Each batch can be restored to a different new directory; the files inside retain their original volume paths. Do not merge outputs by overwriting existing files without comparing hashes.

Restore a checkpoint state

The model weights remain model.safetensors. Large training_state.pt files are split into 8 GiB parts under training_state.pt.parts/. Download that checkpoint's REASSEMBLE.json at the recorded immutable commit, then run:

sh
python restore_backup-v2.py checkpoint REASSEMBLE.json new-checkpoint-directory

The helper checks that parts are contiguous and that the resulting entire file matches the original size and SHA256. Allow room for the reconstructed state plus one temporary part. Do not load a pickle file from an untrusted source merely because its checksum matches a manifest.

Explicit limitations

Credential-like content and opaque Git history are not published. Such entries are marked not_on_hf and are retained by the owner in protected local quarantine with SHA256 manifests. The public archive alone is therefore not a complete restoration of those confidential/opaque entries.

Environment/checkpoint symlinks exposed by the S3 API with empty ETags retain listing metadata only. Exact symlink targets are not claimed preserved; rebuild runtime links from the original runtime manifests. A failed or omitted source payload is never silently relabeled as a symlink.