PR #1927

open

Non-Record: PR #1901 base + LQER Asymmetric + Brotli/Byte-Shuffle Compression

val_bpb

0.8335

Architecture

Transformer

Optimizer

AdaMuon

Artifact Size

18,204 bytes

Training Techniques

Quantization

int6 SDClip

bits: 6

scope: all

INT2/INT4 LQER

bits: null

scope: top-K residual factors

Architecture

SharedMoE

Mixture-of-experts component inherited from PR #1901 base stack.

parameters: null

DualTokenHashSkip

Token hashing / skip-style architectural component inherited from PR #1901 base stack.

parameters: null

Test-Time Training

score-first TTT

parameters: null

Optimizer

AdaMuon

weight_decay: null

momentum: null

other_params: null

Compression

lzma

level: 9

brotli

level: 11

Other

other

Stride-2 byte-shuffle applied before Brotli compression to improve artifact packing efficiency.

parameters: {"stride":2}

LQER asymmetric rank-4 post-quantization correction applied to a Sigma-Delta-quantized stack
Brotli-11 plus stride-2 byte-shuffle replacing LZMA for artifact compression
Theoretical delta-BPB estimate for combining LQER and improved compression
Patched submission provided with LZMA-base85-wrapped train_gpt.py