harmonic-gpt-spherical-kan-128m-maxine

Harmonic GPT Spherical KAN 128M — Maxine

Maxine is an experimental 128.26M-parameter causal byte language model. It starts from the final one-corpus-pass spherical SpectralKAN checkpoint and receives 1,200 steps of benchmark-family instruction tuning over the final two spherical blocks, final normalization, and byte readout head.

Unlike the legacy Max/BENCHMAX checkpoint, Maxine was not trained on the frozen exact public evaluation prompts. It is nevertheless deliberately benchmark-family specialized and should not be presented as a general-purpose or publishable clean evaluation result.

Checkpoint

• Parameters: 128,260,556
• Pretraining step: 122,076
• Pretraining exposure: 8,000,327,497 bytes, one complete corpus pass
• Pretraining validation BPB: 1.115637
• Post-training steps: 1,200
• Updated parameters: final two blocks + final normalization + byte head (14,559,372)
• Post-training mixture: 80% generalized task variants, 10% base replay, 10% clean instruction data
• Checkpoint precision: BF16
• Context window: 2,048 bytes

The model uses a fixed QPSK + Hamming wire encoder, 18 unified blocks, seven heads, eight Funk-Hecke/SpectralKAN channels per head, bounded decayed recurrent state, and projection-free spherical Lohe mean-field coupling. Its hidden width is 1,330.

Native instruction framing

Inference must reproduce the post-training byte format:

Instruction:
{user message}

Response:
{assistant response}

End of response.

The literal End of response. sequence is the trained termination marker and should be stripped by the serving adapter. modeling_spherical_kan.py provides bounded-state token-at-a-time inference through step().

Evaluation warning

The generalized dataset uses synthetic variants of short-form evaluation task families. It excludes the 24 exact prompts in the frozen August 3 public snapshot, but it remains benchmark-directed post-training. The legacy Max checkpoint was directly trained on that snapshot and is an evaluation-contaminated control.

Local deterministic proxy evaluation found that Maxine improved over its guarded step-300 checkpoint but remained below legacy Max on the frozen snapshot. The public eval-site result may differ because it uses a variable task set and judge.

Limitations

This is a small architecture-research model, not a reliable assistant. It may select incorrect short canonical answers, repeat motifs, misunderstand instructions, or state false information. Do not use it for medical, legal, financial, safety-critical, or production decisions.

Released under CC BY-NC 4.0 for research use. Downstream users remain responsible for reviewing source-specific data terms.