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.
- Downloads last month
- 39