Byte Embedder Unit — Baked Winner

C19 + L-BFGS Staged Freeze | Int4 | Tied Mirror
100% Lossless 288 byte int4 24 C19 neuron 40.72% downstream

Architektura

8 bit input {-1, +1} W1 8x24 int4 24 C19 neuron c: 1.5-6.4 | rho: 2.5-5.8 bias: 24 float W2 24x16 int4 16D latent float output MIRROR DECODER 16D → W2^T → 24D → W1^T → 8 bit tied weights (same W, transposed) INPUT HIDDEN LAYER OUTPUT RECONSTRUCTION PARAM BUDGET: W1: 192 int4 (96B) | W2: 384 int4 (192B) | bias: 40 float (160B) | C19 c,rho: 48 float (192B) Weights: 288 byte + 352B float params = 640B total

Specifikaciok

Lossless roundtrip
100%
Downstream char-LM
40.72%
Weight storage
288 B
Compression
8x
Neuronok
24 C19
Optimizer
L-BFGS
Sparsity
27.8%
Train time
~6 min

Suly-eloszlas (int4: -7..+7)

-7-30+3+7
Haranggörbe: 0 korul csucsosodik, 27.8% nulla (sparse). Sulyok -7..+7 kozott, de a tobbseg -3..+3 sav.

24 C19 Neuron — egyenkent

W1 heatmap (8 bit → 24 neuron)

W2 heatmap (24 neuron → 16D)

Live teszt — irj be egy byte-ot (0-255)

= 'A' 8 bit:
→ 24 hidden (C19 output):
→ 16D latent output:
← Mirror decode → reconstructed bits: