MonumentalSystems/clifford-ssm-checkpoints

Clifford-SSM checkpoint recovery archive

Public preservation mirror for the e23-e26 Clifford-SSM training lineages. These files are raw research checkpoints, not packaged inference releases, not a quality ranking, and not a blanket grant to reuse every lineage.

Archive only. Public visibility preserves the research record; it does not mean that every checkpoint is a recommended release or that every underlying dataset has been cleared for unrestricted downstream use.

The archive contains resumable PyTorch checkpoints with model, optimizer, sampler, configuration, step, loss, and BPB state, plus the e23 DGX inference export. Verify files against SHA256SUMS. Local convenience symlinks are recorded in ALIASES.json; the Hub archive stores only the canonical files.

Lineages

• e23_nano_omni_plateau_resume: 119.6M-parameter baseline trained on text-corpus-v1. The saved best.pt alias points to step 23,500; the run log's lowest metric occurred earlier, so "saved best" and "best logged metric" must not be conflated.
• e24_*: 119.6M-parameter Chinchilla continuations and learning-rate/ exposure-governor controls. Several final checkpoints are worse than their early branch-best snapshots.
• e25_quality_80m_chinchilla: completed 84.3M-parameter quality-corpus run.
• e25_quality_fresh_smoke: smoke-test artifact only.
• e26_v2classified_125m_continue: partial classified-corpus continuation; archive-only pending dataset-manifest reconstruction.

Source provenance

• e23 source is represented by private HarmonicRust commits dc464ae and a32645f; the plateau-resume configuration is retained in experiment logs.
• e24 end-state source was recovered and preserved at private commit d66d2bb2f0e249183cf64aee653a15c9e562c995.
• e25 source is represented by private commit 3356a339.
• e26 operative source is represented by private commit 0a866ef5.

The original checkpoints did not embed a source Git SHA. Future checkpoints should record source SHA, dirty-state flag, dependency-lock hash, dataset revision/manifest hash, and launcher arguments.

Safety and licensing

Load only from a trusted immutable revision and use torch.load(..., weights_only=True). These are pickle-based training states.

No public-use license is asserted for this mixed-lineage archive. Dataset provenance differs by lineage and remains under review, especially e24 (license: other) and e26 (local unmanifested corpus). Public inference packages should be emitted separately as Safetensors after lineage-specific review.