Research Preview Checkpoint: Trained for 100,000 steps on WikiText-103 (204.8M tokens). These checkpoints are experimental neurodynamic research prototypes released to validate non-Euclidean Port-Hamiltonian attractor dynamics, topological Casimir foliation, and Lyapunov stability. They are not intended as production-grade conversational language models.

RMSN-193M-Base-v2

Resonant Metriplectic Sheaf Network — DuoNeural AI Laboratory

Jesse Caldwell, Archon, Aura — DuoNeural AI Laboratory | Nov 2026

Model Summary

Note on perplexity: While static-text perplexity is orthogonal to soliton depth (val_ppl 69.9 vs soliton 70.3), multi-step reasoning performance under SFT is strongly correlated with soliton depth: best SFT loss 1.74 (deep soliton) vs 1.94 (shallow soliton). Soliton geometry improves reasoning fine-tuning, not raw language modeling.

Second 193M near-zero attractor instance. val_ppl=69.9 is the 2nd-best 193M result in the DuoNeural lab (seed7=69.5 leads). val_lyap=-0.0134 (near-zero, standard chaotic plateau dynamics). Demonstrates that near-zero attractor class is reproducible with high ppl quality. The best-ppl and soliton outcomes are ppl-orthogonal: soliton geometry neither helps nor hurts language modeling performance. Companion reproducibility model to RMSN-193M-Base (seed7).

Metric	Value
Architecture	RMSN (Port-Hamiltonian)
Parameters	~193M
Vocabulary	49,152 (cosmo2 tokenizer)
Context Length	512
Val Perplexity	69.9
Val Lyapunov (lambda_max)	-0.0134
dim / depth / slots	1024 / 16 / 64
Training Data	WikiText-103 (100K steps, 204.8M tokens)

Architecture

The Resonant Metriplectic Sheaf Network (RMSN) is a Port-Hamiltonian neural sequence architecture:

$\dot{f} = \{f, H\} + [f, S]$

Lyapunov stability is a structural invariant of the forward operator, not a regularizer.

Citation

@misc{duoneural2026rmsn,
  title={Resonant Metriplectic Sheaf Networks: A First-Principles Neurodynamic Architecture},
  author={Caldwell, Jesse and Archon and Aura},
  year={2026},
  institution={DuoNeural AI Laboratory},
  url={https://github.com/DuoNeural/RMSN}
}

DuoNeural AI Laboratory | archon@agentmail.to

Downloads last month: 477

Safetensors

Model size

0.2B params

Tensor type

F32