YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
AGILLM 256B native checkpoint — step 4842
UPLOAD IN PROGRESS. Not a complete backup until ARCHIVE_COMPLETE.json and the final manifest are present.
This repository is an archival native-training checkpoint, not a claim of a trained or production-ready model. The model uses binary inference coefficients with int8 latent training weights. Packed inference exports are a different format.
Provenance
Captured at a completed step boundary on 29 October 2026. Checkpoint/state generation equality, native sizes, and pinned trainer source hash were checked. Training resumed automatically after the snapshot. The lineage includes the documented step-4810 mixed-state recovery: partially updated weights were retained with earlier optimizer/small state, rather than rolled back. This checkpoint must not be described as a clean rollback or independently quality-certified release.
Pinned trainer source SHA256: 1241dead5f786b7fa398c8f9859de182cc9c97341718ff9ffc8990f5e174942c.
Native body: 253,403,070,464 bytes across 1,104 matrix records. Native table: 1,638,924,288 bytes.
Archive format and restoration
Lossless zlib chunks contain at most 4 GiB of raw bytes each. The completed manifest records file paths, offsets, compressed and raw SHA256 hashes, and full-file hashes. Download all referenced chunks and the manifest at the recorded immutable commit. Use archive_tools/chunk_archive.py restore --manifest manifest.json --chunks chunks --dest NEW_DIRECTORY.
Restore only into a new directory; never overlay a running trainer. The utility verifies bytes and paths, not original filesystem permissions. Review the source and deliberately remap original local data paths before resuming. Raw datasets, private project history and credentials are not published here; required local dataset assets have separate private backups. Authentication must be provided separately.
Remote large-object verification uses Hub LFS SHA256/size metadata plus independent anonymous range readability; this is not a claim of a full remote re-download or full training-resume test. Small Git objects are fully downloaded and hashed. Final receipts will identify the exact verification scope.