LEARNS v0.0016: Rhythm — ConsciousNode SoftWorks

Corpus (80 / 10 / 10 Split)

⬇

Drop .txt corpus here

Architecture & Baselines

Architecture (arch)

Settings String

Vocabulary 256

Hidden dim 128

Layers 2

Chunk size 64

Max FP iter (fpIters) 6

Inner FP iter (innerFP) 4

FP ε

Training

Total steps

Auto-checkpoint every 1000 steps

LR start

LR min

Warmup steps

Thermo λ

Thermo mode

Seed (blank/0 = auto; recorded)

R_B mode

Validation fraction (80/10/10 default)

Validate every N steps

Chunk stream (--stream)

Corpus control

φ-asymmetric gate clipping

FP input injection — K_A = F(K_B, a_t)

Frontier reads the present (W_present)

Live sample every N steps

Export name (optional)

Skip optimizer state (smaller file)

Generation

Temperature 0.8

Max tokens

Prompt

Chat

Seed prompt (primes records)

Max response tokens 120

Temperature 0.75

Import

📂

Drop or click to load .learns

Training Progress L12:arch=learns;dim=128

of 4000 steps — awaiting corpus

Loss — Total (accent) · Thermodynamic (dim)

Ternary Weight Balance — zeros should grow

−1

33%

34%

33%

Diagnostics

Loss

—

FP Iter avg

—

FP Residual

—

FP Agree

—

R_A Entropy

—

R_B Entropy

—

F Entropy

—

A sym iter

—

A act

—

B act

—

Skew Cos(A,B)

—

LN Spikes

R_A bias ↑+

—

R_B bias ↓−

—

Divergence

—

Eff ctx

—

Val loss (nats)

—

Val bits/byte

—

Val ppl

—

Unigram base

—

Live Sample — updated periodically

Awaiting first training steps...

Switchboard — Every Component a Switch Defaults reproduce v0.0011 bit-for-bit

1. Relational Embedding It's All Relative

relEmbed Default: on

Token meaning is embedding plus Q(W_rel)·R_combined_prev. Off = embedding alone.

2. Temporal Turing Core The Electromagnetic Node

coreA Default: on

Component A inner fixed-point pattern recognizer over recent suffix_sig.

coreB Default: on

Component B reads current helix h and R_combined_prev.

innerFP Default: 4

Maximum inner fixed-point iterations for Component A (2–8).

3. Dyadic Fixed Point K-H Fixed Point Theorem

fpIters Default: 6

Jacobi iterations for K_A and K_B (0 = single feed-forward pass).

fpInject Default: on (v0.0014)

DEQ-style input injection K_A = F_A(K_B, a_t), K_B = F_B(K_A, b_t).

fpNorm Default: rescale (v0.0014)

Jacobi normalization: rescale/none (contractive, σ(W_F)≤0.9) vs layernorm (v0.0011 legacy).

4. Record / Frontier Split It's Time To Toss The Dice

recordA Default: on

Forward record R_A accumulating K_A and cross-coupling from R_B.

recordB Default: on

Partner record R_B and auxiliary reverse-LM scout training.

rbMode Default: causal

How R_B is updated or seeded (causal carry, rolling backward lookback, prevrev, or legacy).

lookback Default: 16

Look-back window N for rbMode=lookback (runs R_B backwards over x_t..x_{t-N+1}).

scoutScale Default: 0 (v0.0015)

Reverse-chunk auxiliary gradient weight (0 = default; 0.5 = reverse-LM scout active).

scoutHead Default: shared

Readout head used by reverse scout when scoutScale > 0 (shared vs separate head).

frontierPresent Default: on

Frontier reads working state via Q(W_present) in addition to Q(W_frontier)·R.

frontierHelix Default: on

Include helical recurrence σ(ρ)⊙G(θ)h in frontier readout vs instantaneous P_t only.

5. Helical State The Drill of Entropy

helix Default: on

Master switch for helical recurrence σ(ρ)⊙G(θ)h_prev across tokens.

helixRotation Default: learned

Learned Givens rotation angles θ per layer vs identity rotation.

helixRetention Default: learned

Per-dimension retention σ(ρ) vs fixed σ(2.2)≈0.9 vs un-decayed (v0.0009).

helixState Default: shared

Shared d-dim helical state vs independent per-layer spirals hs[l] (L×d state floats).

6. Thermodynamic Loss The Past Is Still The Future

thermo Default: hinge

Penalize record entropy drop over chunk (hinge: λ·max(0,−ΔS); linear: −λ·ΔS).

lambda Default: 0.05

Thermodynamic penalty coefficient λ.

7. Numerics & Gradient Clipping BitNet b1.58 STE & Spike Isolation

weights Default: ternary

Ternary (−1, 0, +1 × mean|W| with exact STE) vs full-precision float32.

clipMode Default: global

Global L2 norm clip vs per-parameter-group clip (isolates LayerNorm-beta spikes).

phiClip Default: on (inert)

Post-norm gate elementwise clamp (±φ / ±1/φ). Inert after norm clip ≤ 1.

ema Default: off

Evaluate exponential moving average (β=0.99) of float shadow weights quantized at eval time.

Record entropy during generation

// Generation output will appear here