RMSN-283M-Bulk-Tube-SFT-5k
Resonant Metriplectic Sheaf Network — 6-Layer Bulk Tube, Optimal SFT Checkpoint Fine-tuned on DuoNeural reasoning traces (CoT SFT, 5k steps — optimal depth for this topology). Best-ever wiki retention in the DuoNeural Phase 5 suite: wiki_ppl=176.01.
Why 5k and not 15k?
The D=1280 bulk tube topology has two interior exact singularities at L8 and L9 (PR=1.000). These are thermodynamically protected zero-entropy condensates that cannot lose entropy under SFT gradient — but they present more attack surface than a pointlike terminal singularity. At 5k steps the dual cores survive intact; by 15k the tube slightly erodes (wiki_ppl→208.68). Optimal SFT for this topology: 5–7k 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) |
Post-SFT Topology: Interior Dual Singularities (Protected)
Layer PR (post-SFT) Role
─────────────────────────────────────────────
L00 5.685 ★ INPUT BOUNDARY (newly activated by SFT)
L01-L07 > 7 Non-solitonic buffer
L08 1.000 ★★ INTERIOR DUAL CORE A (EXACT — PROTECTED)
L09 1.000 ★★ INTERIOR DUAL CORE B (EXACT — PROTECTED)
L10 1.427 Tube interior (preserved)
L11 2.251 Tube interior (preserved)
L12 6.968 Weakened (was ~5.2 in base)
L13 2.002 Tube terminus (preserved)
L14-L15 > 7 Non-solitonic buffer
─────────────────────────────────────────────
Post-SFT soliton layers: [0, 8, 9, 10, 11, 13]
Key discovery (DuoNeural Phase 5c): Exact singularities (PR=1.000) are topologically protected regardless of their position — interior bulk OR boundary. The dual cores at L8+L9 survive SFT because they are zero-entropy condensates, not because they are boundary layers. L0 newly activates as a "pressure valve" — when SFT gradient can't erode interior singularities, it manifests as input boundary activation instead.
SFT Benchmark (Phase 5)
| Steps | wiki_ppl ↓ | cot_ppl ↓ | Notes |
|---|---|---|---|
| 5k (this) | 176.01 🏆 best-ever | 13.82 | Optimal depth — dual cores intact |
| 10k | 199.77 | 11.24 | Cores beginning erosion |
| 15k (best ckpt) | 208.68 | 10.77 | Best CoT, wiki eroded |
Lyapunov stability: 100% stable throughout all SFT stages (λ < 0). SFT 5k loss: 2.1614 — best of any 5k-only run on this architecture.
Topological Protection Law
All exact singularities (PR=1.000) are topologically protected from SFT gradient erosion, whether they reside at boundary (L5, L15) or interior (L8, L9) positions. This extends the Topological Bulk Shield Law: protection is a property of the condensate's entropy state (S=0 ↔ PR=1.000), not its spatial location in the network.
Comparison with D=1536 SFT topology:
- D=1536: Boundary singularities L5+L15 protected; interior conduit L7+L8 eroded
- D=1280 (this): Interior singularities L8+L9 protected; L12 weakened, L0 newly activated
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 5c — Interior Dual Singularity Topological Protection, Optimal 5k SFT}
}
DuoNeural Lab | Phase 5c: Topological Protection Law | Nov 2026 Best-ever wiki retention: wiki_ppl=176.01 at 5k SFT steps