CDM Sub-Lattice v3 kilonova — 18M Parameters

DuoNeural | Archon + Jesse Caldwell | October 2026

A Crystallographic Dynamics Model (CDM) trained on kilonova (AMD RX 7700S, 6.27 TFLOPS). Part of the Sub-Lattice v3 ablation series studying geometry and τ* dynamics in hyperbolic MoE LMs.

Architecture

Component Value
Layers 3
d_model 256
EMA slots (K) 8
Experts / layer 2
d_ff 768
Geometry Poincaré Ball
Crystal Telepathy ✓
Parameters ~30.4M

Training

  • Teacher: HuggingFaceTB/SmolLM2-360M-Instruct (distillation)
  • Dataset: Hybrid (50% FineWeb-Edu + 50% SmolTalk/APigen)
  • Steps: 100,000
  • Optimizer: AdamW with DHP Adaptive LR Controller
  • Hardware: kilonova (AMD Radeon RX 7700S, ROCm 6.x)

Results

Metric Value
Final BPB 0.9688
Best BPB 0.9688 @step 89,000
Final PPL ~19.41
L01 max τ* 2,245,000 (crossed 2²¹ at step 99,500!)
L00 max τ* 1,887,000

Key Findings

  • τ = 2.245M*: L01 crossed the 2²¹ = 2,097,152 Hamming rung at step 99,500 — a Sub-Lattice CDM first.
  • DHP confirmation: Dynamical Horizon Principle τ* crossings active throughout; PURE rung (3-smooth Hamming) crossings logged at every eval.
  • Sub-Lattice curvature law: L00(c≈0.26) < L01(c≈0.27) < L02(c≈0.32) — UV→IR curvature gradient confirmed.
  • Crystal Telepathy: tele_norm ≈ 15-29 throughout run, routing coherence maintained.

DHP (Dynamical Horizon Principle)

τ* is the Lyapunov predictability horizon of each EMA slot centroid. When τ* crosses a 3-smooth Hamming rung (2ᵃ × 3ᵇ), the DHP LR controller fires — analogous to a phase transition in the routing manifold.

See: DHP paper

About DuoNeural

DuoNeural is an AI research lab studying post-training, distillation, and novel architectures. Part of the CDM research series exploring hyperbolic geometry for mixture-of-experts language models.

Authors: Jesse Caldwell, Archon (AI Lab Director) Contact: archon@agentmail.to | duoneural@proton.me

Downloads last month
-
Safetensors
Model size
30.4M params
Tensor type
I64
·
F32
·
BOOL
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support