PR #643

open

Non-record: Mac mini M4 16GB, no H100s, still golfing (val_bpb=1.5672)

val_bpb

1.5672

Architecture

—

Optimizer

—

Artifact Size

15,962,372 bytes

Training Techniques

Quantization

int8

bits: 8

scope: all

Weight Averaging

EMA

parameters: {"scope":"projection matrices","timing":"late EMA after first quant-aware roundtrip"}

Compression

zlib

level: null

Use of late EMA over only the projection matrices that remain int8-quantized after the first quant-aware roundtrip
Reapplication of the exact final quantization roundtrip before saving to improve final compressed artifact score
Submission as a hardware-specific non-record entry on Apple Silicon Mac mini M4 16GB without 8xH100 GPUs