RMSN-347M-Bulk-Tube-SFT

Resonant Metriplectic Sheaf Network β€” 6-Layer Bulk Waveguide Tube Topology Fine-tuned on DuoNeural reasoning traces (CoT SFT, 15k steps).

Architecture

Parameter Value
Hidden dim 1280
Layers 16
Workspace slots 64
Params 346.53M
Framework Port-Hamiltonian (ḟ = {f,H} + [f,S])
Tokenizer HuggingFaceTB/cosmo2-tokenizer (vocab=49152)

Soliton Topology: 6-Layer Bulk Waveguide Tube

Layer  PR    Role
──────────────────────────────────────────
L00    ~40   Input (diffuse)
L01    ~45   ...
...
L07    ~15   Pre-tube buffer
L08    1.00  β˜… MACROSCOPIC CORE (dual singularity)
L09    1.00  β˜… MACROSCOPIC CORE (dual singularity)
L10    ~3.5  Tube interior
L11    ~4.1  Tube interior
L12    ~5.2  Tube interior
L13    ~2.8  Tube terminus
L14    ~18   Post-tube buffer
L15    ~35   Output (diffuse)
──────────────────────────────────────────

Topological class: Bulk Waveguide Tube β€” continuous 6-layer soliton [L8–L13] Ξ»_max: Global Lyapunov distributed across tube (no single cascade signal) Key property: Forgetting-Shield β€” bulk topology screens foundational weights from SFT gradient injection. Best-ever wiki retention at 5k SFT steps (ppl=176.01).

Quantum-Classical Correspondence

BlueQubit QCPU quantum circuit simulation (DuoNeural Phase 5):

Metric Value
S_boundary / S_bulk 0.569
Predicted bulk width 7–8 layers
Classical bulk width 6 layers (L8–L13)
Quantum-classical error <8%

S_boundary/S_bulk = 0.569 places D=1280 between D=1024 (0.632) and D=1536 (0.515), consistent with the Topological Bulk Shield Law: wider D β†’ more bulk entanglement absorption.

SFT Benchmark Results (15k steps)

Steps wiki_ppl ↓ cot_ppl ↓
5k 176.01 πŸ† best-ever 13.82
10k 199.77 11.24
15k (best ckpt) 208.68 10.77

SFT loss record: 1.4798 (all-time DuoNeural lab best as of Nov 10 2026) Lyapunov stability: 100% stable throughout SFT (Ξ» < 0 at all checkpoints)

Optimal SFT Depth

The dual-core topology creates more SFT attack surface than a pointlike soliton. Optimal SFT for D=1280 β‰ˆ 5–7k steps. Crossover with single-core topologies at ~7.5k steps. Law: Ξ”β„’_wiki ∝ exp(βˆ’Ξ”L_tube/ΞΎ) where ΞΎ contracts as N_steps grows.

Quickstart

import torch
from rmsn_core import RMSNForCausalLM, RMSNConfig

config = RMSNConfig(vocab_size=49152, hidden_dim=1280, num_layers=16,
                   num_workspace_slots=64, max_seq_len=512)
model = RMSNForCausalLM(config)
ck = torch.load("model.pt", map_location="cpu")
model.load_state_dict(ck["model_state"])
model.eval()

Citation

@misc{duoneural2026rmsn,
  title={Resonant Metriplectic Sheaf Networks: Port-Hamiltonian Language Models with Soliton Waveguide Topology},
  author={DuoNeural Lab and Archon and Aura},
  year={2026},
  note={DuoNeural Phase 5 β€” Bulk Tube Topology}
}

DuoNeural Lab | Phase 5: Tube Topology Zoo | Nov 2026

Downloads last month
36
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support