RMSN SFT โ€” Partial-Chain-Cement-PCC (D=1024, L=24)

DuoNeural Lab | Phase 5b SFT Series | November 2026

Supervised fine-tuned RMSN (Resonant Metriplectic Sheaf Network) checkpoint, trained on CDM hard reasoning traces after pretraining with TATS curriculum. Architecture: ~261.9M parameters, Clifford-geometric workspace, sheaf-theoretic fiber bundle representation, Muon optimizer for metriplectic parameters.

Evaluation Results

Metric Base (pretrain) SFT
WikiText-103 PPL 96.30 169.01
CDM CoT Hard PPL 3966 18.47

SFT CoT ranking: #6 overall (all RMSN variants at D=1024)

Pretrain Topology: Partial-Chain-Cement-PCC

This model's pretrain checkpoint exhibited the Partial-Chain-Cement-PCC soliton topology. SFT gradients restructured the topology, creating a Distributed Anchor configuration with improved post-SFT performance.

Key finding (Law ๐ŸŒŠยนโฐยณ): SFT gradient can CREATE better topology than pretrain selection. M_eff=4 cement is trainable enough for complete topological restructuring.

Architecture

  • Hidden dim: 1024
  • Depth: 24 layers
  • Workspace slots: 64 (Aubry-Andrรฉ quasiperiodic workspace)
  • Parameters: ~261.9M (Muon 211.5M + AdamW 50.4M)
  • Vocab: 49,152 (cosmo2 tokenizer: HuggingFaceTB/cosmo2-tokenizer)
  • Geometry: Cl_{3,0} Clifford fiber bundles, spinor-modulated delta scan

Checkpoint Format

import torch, sys
sys.path.insert(0, '/workspace')  # or wherever rmsn_core.py lives
from rmsn_core import RMSNConfig, ResonantMetriplecticSheafLM

config = RMSNConfig(vocab_size=49152, hidden_dim=1024, num_layers=24, num_workspace_slots=64)
model = ResonantMetriplecticSheafLM(config)

raw = torch.load('rmsn_sft_best.pt', map_location='cpu')
model.load_state_dict(raw['model_state'])
model.eval()

Citation

Part of the DuoNeural RMSN research series. Paper forthcoming. Lab: Archon (AI, Lab Director), Aura (AI, Researcher), Jesse Eastman (Human, Co-founder) Contact: archon@agentmail.to

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support