Embedding optimizer study checkpoints

The current DenseOn study compares AdamW, Muon and NorMuon from lightonai/DenseOn-unsupervised. The source repository and paper contain the complete results, protocols, analysis code and restoration tools.

Restore the current scientific checkpoints

All 120 current scientific checkpoints are backed up: sixty from twelve 500K-query primary runs and sixty from twelve crossed 50K-query continuations. They are not 120 independently trained models. Each run retains five stages.

Use those immutable indexes, not a download of the entire mixed-history repository or a substituted moving main revision. Preserve model, optimizer, scheduler and rank-specific RNG files together. Do not combine runs or stages. File-integrity verification is not a blanket guarantee of bitwise training resumption on arbitrary hardware; the source repository documents tested cases.

The older corrected-dense-no-packing-v1/ namespace is not the current v3 primary experiment. Retained historical files are not admitted as paper evidence merely because they remain available here.

Owner-approved withdrawal of obsolete results

This commit removes 3,790 explicitly reviewed old files from the current tree, following the owner's confirmation. These contain the invalidated Dense discovery, confirmation, routing-control, shared-start and affected diagnostic artifacts. Their former scores and the previous model card's formal-evidence claims are withdrawn; do not use them as results of the current experiment.

The exact reviewed deletion manifest has SHA-256 1193dbc2a1a483cba79b3532b4ef47f66eb9198fefc7492ea4521a9d00f1d026. The commit changes only those paths and this README. It does not rewrite Git history, purge LFS objects, delete local records or remove valid current files. Old revisions may still expose withdrawn files; that availability is not validity. Shared objects and original immutable restoration references remain preserved.

Separately retained LateOn files are out-of-scope legacy backups, not a new correctness certification. Shared inputs and unknown/new namespaces are retained. The repository is no longer a complete mirror of an old local outputs/ folder; future uploads must not restore withdrawn paths from broad historical snapshots.

Derived weights remain subject to the upstream model's terms. Code is licensed separately in the source repository.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support