← PR #2155
PR #2158→

PR #2157

open

Record candidate: PR #1797 + AWQ-lite top3 + LQER 60k on b180-tlr56 — val_bpb 1.06043 (seed=0)

by vimetoView on GitHub
val_bpb
1.0604
Architecture
Transformer
Optimizer
—
Artifact Size
15,982,182 bytes

Training Techniques

Quantization
GPTQ-lite
bits: null
scope: block weights
mixed int7/int8
bits: 8
scope: embeddings and selected groups
int8
bits: 8
scope: top-3 salient 64-col groups
Test-Time Training
LoRA TTT
parameters: {"rank":56}
Regularization
dropout
parameters: {"drop_M":true}
logit softcap
parameters: null
Other
other
LQER budget reduction to 60k bytes, using cap margin freed by AWQ-lite promotions
parameters: {"budget_bytes":60000}
other
SparseAttnGate lineage contribution
parameters: null

Novel Contributions

  • AWQ-lite mixed-precision GPTQ stacked on top of the b180-tlr56 lineage
  • Top-3 saliency 64-column groups promoted to INT8
  • Reduced LQER budget to 60k using freed artifact margin
  • Drop-M LoRA / TTT rank-56 continuation of the PR #1935 lineage
  • Per-group lrzip artifact serialization with lossless roundtrip verification