You can deploy this model and start earning money today!

MonumentalSystems/
harmonic-gpt-checkpoints

Name: harmonic-gpt-checkpoints
Availability: InStock

geometric-deep-learning

Views

No views yet

HarmonicGPT Research Checkpoints

Public, provenance-preserving checkpoints from HarmonicGPT experiments. These are research artifacts rather than a packaged Transformers model: use the matching HarmonicGPT architecture and training code to load them.

Archive Inventory

The repository preserves 113 checkpoints in 18 top-level archive groups, spanning 30 named runs and subruns. The inventory includes full trajectories when intermediate states are useful for reproducing non-monotonic learning behavior, plus seed sets, curriculum-stage outputs, and precision/gradient checkpointing comparisons.

Run family	Checkpoints	Recorded step span	Model parameters
`manifold_distill_checkpoints`	7	500-3,000	179.6M
`manifold_wave_native_curriculum`	8	staged outputs	25.0M
`manifold_wave_native_curriculum_v2`	5	staged outputs	25.0M
`paper_a_s42`, `paper_a_s123`, `paper_a_s456`	9	1,847-1,999	19.2M
`rotor_25m_annealed`	3	63,915-64,915	25.5M
`rotor_25m_chinchilla`	15	5,000-61,915	25.5M
`rotor_25m_salon_cot`	3	2,500-3,000	25.5M
`rotor_25m_synth_ft`	3	1,000-2,000	25.5M
`rotor_projfree_quality`	9	5,000-31,180	12.9M
`shifts_compare`	9	472-499	19.2M
`unified_25m_adult`	7	5,000-23,246	19.2M
`unified_25m_chinchilla_v2`	2	3,500-3,563	19.2M
`unified_25m_quality`	3	5,000-8,000	19.2M
`unified_coulomb_test`	3	1,984-1,999	19.2M
`wavenative_25m_continued`	4	2,000-6,000	25.1M
`wavenative_25m_continued_fresh`	3	3,891-5,000	25.5M
`harmonic-gpt-codex/bench_runs`	7	four 3060 precision/GC subruns	See catalog
`harmonic-gpt-codex/runs`	13	five corpus/phrasebook subruns	See catalog

The root checkpoint-catalog.json is the machine-readable index across all families. Each raw checkpoint is also covered by the root SHA256SUMS file.

RotorNative 25M Chinchilla Run

rotor_25m_chinchilla/ contains the complete 5,000-step checkpoint trajectory for a 25.5M-parameter byte-level RotorNative language model, plus its best and Chinchilla-budget endpoints.

Artifact	Step	Observed BPB	Purpose
`best.pt`	11,099	0.437599	Best observed training-stream BPB
`checkpoint_5000.pt` ... `checkpoint_60000.pt`	5,000-step intervals	See catalog	Training trajectory and intermediate comparisons
`final.pt`	61,915	0.521 at the final logged step	Chinchilla-budget endpoint
`latest.pt`	61,915	Same model weights as `final.pt`	Auto-resume alias retained in the recovery archive

The run used a developmental 500 MB byte-token curriculum, sequence length 256, effective batch size 32, RotationalAdamW, 20 layers, and 40 harmonics. The Chinchilla budget was calculated as 20 tokens per parameter and produced 61,916 optimizer steps. The full run took about 10.4 hours at roughly 13.6k tokens/s.

The trajectory is intentionally retained because its BPB was non-monotonic: it reached its best value early, degraded through the middle of the run, and then partially recovered near the endpoint. That makes the intermediate checkpoints scientifically useful rather than disposable periodic saves.

Provenance Notes

checkpoint-catalog.json records file SHA-256, canonical model-state SHA-256, tensor counts, parameter counts, run scalars, and checkpoint configuration.
SHA256SUMS verifies every raw artifact.
training.log is the original run log.
final.pt and latest.pt contain identical model-state hashes. They are separate PyTorch containers because they were written independently.
The save routine wrote best_bpb into the final checkpoint's bpb field. Therefore the final checkpoint metadata says 0.437599, while the original log reports 0.521 at step 61,915. The log is authoritative for endpoint BPB.
The run occurred on 2026-04-19. Its exact dirty-worktree source revision was not embedded in the checkpoint. The closest earlier repository commit was cb117fe72fd0145ac475e4b1782942492fbd4255; the training script was later committed in 4cc85a81239488fe4f8623dc0cc1d8953f421918. Do not treat either as an exact source snapshot without further reconstruction.

Two checksum-verified recovery copies are also retained outside Hugging Face on independent local storage targets (spin_storage and tank). The archive copies include raw provenance and are the recovery authority; this Hub repo is the public, discoverable mirror.

Loading

PyTorch checkpoints include model_state, optimizer_state, step, cursor, BPB fields, and a compact architecture config. Verify SHA256SUMS first and load only artifacts from a trusted revision.

python

import torch

checkpoint = torch.load("best.pt", map_location="cpu", weights_only=True)
model.load_state_dict(checkpoint["model_state"])

No inference API, safety evaluation, or production suitability is claimed. No standalone license has yet been declared for these checkpoint artifacts.

MonumentalSystems/harmonic-gpt-checkpoints