QLLM TRAINER — First Quaternary Language Model

QLLM·TRAINER

First Quaternary Language Model · Native 4-State Training

0.0000NET φ

—LOSS

—PERPLX

0STEP

0PARAMS

QUAT·4

Training Mode

4QUAT
G T A C

128CHAR
ASCII

133WORD
QuatOS

Corpus Sources

DNA SEQUENCES

2.2M native GTAC tokens

QUATOS DOCS

740KB domain knowledge

PHI STREAM

56,852 phi measurements

CONVERSATIONS

Live QuatOS sessions

Architecture

d_model64

layers3

seq_len32

batch4

learn rate.003

φ_reg.10

▶ START TRAINING

Gate Bias (L2L)

25%

strategy: exploratory

ARCHITECTURE

ATTENTION MAPS

CORPUS PREVIEW

GENERATION

Loading corpus...

QLLM GENERATION

Start training, then use prompts to generate text. In QUAT mode, the model generates GTAC sequences. In CHAR/WORD mode, it generates QuatOS language.

GENERATE

STEP 0 LR — φ — δ — MODE QUAT·4 QuatOS offline

Loss Curve

LOSS—

Phi Convergence

φ NET—

Perplexity

PPL—

Model Metrics

VOCAB SIZE4

PARAMS (uint8)—

COMPRESSION4×

BEST LOSS—

BEST φ DELTA—

CORPUS TOKENS—

STRATEGYEXPLORATORY

SYSTEM φ—

DNA Weight Strip

Layer Phi Health

Event Log